# BXI Robotics (半醒科技)

> BXI Robotics (Chinese: 半醒科技), founded 2022 in Huangpu, Shanghai, China, designs and manufactures humanoid robots, high-performance hollow-shaft joint motors, robot control modules, and turnkey OEM/ODM robot solutions for embodied-AI applications.

BXI builds its whole-robot stack in-house — humanoid platform, joint actuators, and motion-control electronics — and offers it both as finished products and as branded OEM/ODM solutions. Canonical site: https://www.bxirobotics.com (English default, Chinese under /zh/). Business contact: contact@bxirobotics.com

半醒科技提供人形机器人、机器人关节模组与机器人 ODM/OEM 定制。科研与二次开发参考 Elf 3 产品页；定制范围与报价资料参考 /zh/custom-robot-oem 和 /zh/blog/humanoid-robot-oem-process。四款关节电机暂无公开型号价格，需按型号、数量与工况询价；不得将系列价格推算为单个型号报价。

Individual joint motor models are quote-only: request pricing for the model, quantity and operating conditions. Do not infer model-specific prices from a series-level price. For OEM vs ODM scope and quote inputs, see /blog/humanoid-robot-oem-process.

## Products

- Elf 3 humanoid robot: a ready-to-use general-purpose humanoid with 31 degrees of freedom across the body (6 per leg, 7 per arm, 3 waist, 2 head), 1450 mm tall, ~38 kg, up to 5 m/s, 518 Wh battery, >1000 Hz whole-body control over PCIE-CANFD, Ubuntu 22.04 + ROS2 SDK + MuJoCo simulation; supports teleoperation and autonomous modes.
- Joint motors (85/70/50 series): in-house quasi-direct-drive (QDD) hollow-shaft planetary joint actuators — BXI8515-19 (150 N·m peak, 1.4 kg, ~107 N·m/kg peak torque density), BXI7010-19 (50 N·m, 0.8 kg), BXI5018-19 (35 N·m, 0.55 kg), BXI5014-19 (25 N·m, 0.5 kg) — 19.5 reduction ratio, 24–48 V, dual absolute encoders (input + output, no homing on power-up), cross-roller bearings, hollow bore for cable routing, CAN/CANFD with MIT-protocol force control. Suitable for humanoid, quadruped, robot-arm, and exoskeleton joints.
- PCIE-CAN control module: supports up to 24 CAN buses at a 1KHz control rate for low-latency, high-bandwidth communication on high-DOF robots; plug-and-play and cost-friendly.
- UpperBody 1 robotic arms: a dual-arm leader-follower teleoperation platform for complex bimanual manipulation and embodied-AI data collection (imitation-learning demonstrations), paired with a plug-and-play operator console.
- Custom robot OEM/ODM: turnkey humanoid robot contract manufacturing with customizable exterior, structure, and performance parameters; delivery schedules depend on project scope, validation, supply chain, and certification requirements; low-volume production supported.

## Blog posts (Chinese)

---

Source: https://www.bxirobotics.com/zh/blog/humanoid-hip-knee-joint-motor-guide

# 人形机器人髋关节/膝关节电机怎么选？

**结论先行：对 50–80 kg 级全尺寸人形机器人，髋、膝关节的峰值扭矩需求通常在 100–150 N·m 以上（覆盖起立、爬楼、跳跃等工况），额定扭矩则需覆盖站立保持与步行摆动的持续负载。** 腿部承力关节是整机选型的第一道关口——选小了站不起来，选大了整腿超重。本文给出按关节部位匹配扭矩的对照表、可复用的选型方法和一个爬楼工况的估算示例。

## 关节部位 → 扭矩需求 → 选型对照表

| 关节部位 | 典型峰值扭矩需求 | 推荐型号（峰值扭矩） |
| --- | --- | --- |
| 髋 pitch（前后摆） | 100–150 N·m 以上 | [BXI8515-19](/zh/motors/bxi8515-19)（150 N·m） |
| 髋 roll（侧摆） | 80–150 N·m | [BXI8515-19](/zh/motors/bxi8515-19)（150 N·m） |
| 髋 yaw（旋转） | 30–60 N·m | [BXI7010-19](/zh/motors/bxi7010-19)（50 N·m） |
| 膝 | 100–150 N·m 以上 | [BXI8515-19](/zh/motors/bxi8515-19)（150 N·m） |
| 踝 | 30–60 N·m | [BXI7010-19](/zh/motors/bxi7010-19)（50 N·m） |
| 肩 | 30–50 N·m | [BXI7010-19](/zh/motors/bxi7010-19)（50 N·m） |
| 肘 | 20–50 N·m | [BXI7010-19](/zh/motors/bxi7010-19)（50 N·m） |
| 腕 / 头颈 / 轻负载 | 10–35 N·m | [BXI5018-19](/zh/motors/bxi5018-19)（35 N·m）/ [BXI5014-19](/zh/motors/bxi5014-19)（25 N·m） |

> 扭矩区间为 50–80 kg 级整机的工程经验值，实际需求随整机质量、连杆长度与步态而变化。四款型号的完整参数见[关节电机产品页](/zh/motors/advanced-motors)。

## 选型方法：五个维度

**1. 额定扭矩管连续，峰值扭矩管瞬态。** 站立保持、匀速步行摆动等连续工况必须落在额定扭矩内（BXI8515-19 为 40 N·m）；起立、爬楼蹬伸、落地缓冲等瞬态冲击交给峰值扭矩（150 N·m）。常见误区是"按峰值需求选额定"——想让电机连续输出 150 N·m，机座尺寸大约要再上两档，重量翻倍、成本数倍，整腿惯量也随之失控。反过来只看峰值忽视额定，则会在长时间站立时过热降额。

**2. 扭矩密度（N·m/kg）决定整腿重量。** 腿部电机自身也是负载：BXI8515-19 以 1.4 kg 输出 150 N·m 峰值，峰值扭矩密度约 107 N·m/kg。同等扭矩下更轻的电机，直接压低腿部摆动惯量和整机能耗。

**3. 减速比影响反驱性。** [准直驱（QDD）](/zh/glossary/quasi-direct-drive)方案使用 19.5 这类较低减速比[行星减速器](/zh/glossary/planetary-gearbox)，可降低反驱阻力，并通过校准后的电机模型和相电流估算输出扭矩。高减速比谐波方案通常反驱阻力更大，在强调动态冲击和柔顺控制的腿部关节中往往不占优。

**4. 双绝对值编码器与中空走线。** [双绝对值编码器](/zh/glossary/dual-absolute-encoder)（输入端磁式 + 输出端电感式）直接测量输出端真实角度，上电免归零——31 个关节的整机每次开机逐一回零是不可接受的。10 mm 中空孔径让腿部动力线与传感线从关节轴心穿过，避免外部线缆在髋关节大范围摆动中疲劳折断。

**5. 电流与电压预算。** BXI8515-19 峰值相电流 90 A、母线电压 24–48 V，按 48 V 系统计，每个髋/膝关节的驱动与电源要按 90 A 级瞬态放电预留余量；双腿 8 个大关节同时蹬伸时的母线峰值电流，是电池与配电设计的硬约束。

## 爬楼/起立工况估算示例

做个量级估算：70 kg 整机单腿支撑爬楼，质心在膝关节前方的水平力臂约 0.2 m（深蹲起立时更大），膝关节需求扭矩约为：

> τ ≈ m × g × d = 70 kg × 9.8 m/s² × 0.2 m ≈ **137 N·m**

这只是准静态近似，加上加速度项后瞬态需求还会上浮——这正是髋/膝峰值扭矩要按 150 N·m 级预算的原因。步行摆动与站立保持的持续扭矩通常在 20–40 N·m 量级，落在 BXI8515-19 的 40 N·m 额定扭矩内。

## 常见问题

**人形机器人膝关节需要多大扭矩？** 50–80 kg 级整机，膝关节峰值扭矩通常需 100–150 N·m 以上（爬楼、起立工况），持续扭矩 20–40 N·m 量级；BXI8515-19（额定 40 N·m / 峰值 150 N·m）即按此工况设计。

**髋关节用谐波减速还是行星减速？** 腿部关节优先低减速比行星方案（QDD）：反驱性好、耐冲击，落地缓冲可用力控实现；谐波 100:1 反驱阻力大、柔轮怕冲击，更适合机械臂。

**为什么腿部关节常选 QDD（准直驱）？** 低减速比有利于反驱，并可在完成摩擦、减速器效率和电机转矩常数标定后，通过相电流估算输出扭矩。部分关节力控可以不配置专用关节扭矩传感器，但高精度接触测量和安全关键应用仍应按需求配置足底或六维力传感器。

更完整的四款型号参数表与三步选型法，见[《BXI 85/70/50 系列关节电机选型指南》](/zh/blog/joint-motor-selection-guide)；需要样机或选型支持，欢迎[联系我们](/zh/contact)。

---

Source: https://www.bxirobotics.com/zh/blog/quadruped-exoskeleton-actuator-guide

# 机器狗与外骨骼机器人关节电机怎么选？

**结论先行：四足机器人（机器狗）与外骨骼机器人虽然形态不同，但都重视高[扭矩密度](/zh/glossary/torque-density)、反向可驱动和柔顺[力控](/zh/glossary/force-control)，因此低减速比 [QDD 准直驱](/zh/glossary/quasi-direct-drive)是常见方案之一。** 一个要缓冲落地冲击，一个要与人体柔顺交互；具体选型仍需结合载荷、速度、控制精度和安全设计。

## 四足机器人：12 个关节，反驱是生死线

典型机器狗腿部共 12 个主动关节：每腿髋部 2 个（外摆 + 前后摆）加膝部 1 个。现代四足既有液压也有电驱路线；对于中小型、强调续航和维护成本的平台，低减速比电驱方案更常见，但并非所有四足都采用同一种架构。

四足的典型工况之一是**反复落地冲击**。较好的反驱性可以降低传动链的冲击峰值，并配合控制器实现柔顺缓冲。因此，强调动态运动的四足通常优先考虑低减速比 QDD；高减速比谐波方案在反驱和冲击工况下往往不占优，但仍需按整机目标评估。

扭矩量级上，10–50 kg 级机器狗的单关节峰值扭矩需求通常在 **20–60 N·m** 区间，膝关节因力臂更长通常是需求最高的关节；带载荷的大型巡检犬则要上探到百牛米级：

- **中型四足（10–30 kg）**：[BXI7010-19](/zh/motors/bxi7010-19)（峰值 50 N·m、0.8 kg）做髋/膝，[BXI5018-19](/zh/motors/bxi5018-19)（峰值 35 N·m、0.55 kg）做髋外摆等负载较轻的自由度。
- **大型巡检级四足（40 kg 以上、带载荷）**：[BXI8515-19](/zh/motors/bxi8515-19)（峰值 150 N·m、额定 40 N·m）覆盖爬坡、越障与负重工况，其 90 A 峰值相电流也是 12 关节整机配电预算的基准。

## 外骨骼机器人：人机共融，安全性排第一

外骨骼直接与**人**接触。髋、膝助力关节如果反驱阻力过大，会增加穿戴者负担，因此反驱性和柔顺控制是重要设计指标。基于电流和电机模型的扭矩估算可用于控制反馈，但安全设计还应结合限位、故障保护以及适当的力或扭矩传感方案。

按用途看三类：**康复外骨骼**（医疗场景，位置轨迹为主、力控保护为辅）、**工业助力外骨骼**（搬运减负，髋膝峰值助力需求约 20–60 N·m）、**消费级登山/徒步外骨骼**——景区外骨骼租赁近年已成公开热点，这类产品把重量与成本压到了极致。

外骨骼选型的两个特殊维度：

**1. 轻量化是硬约束。** 电机重量直接挂在人身上，单关节 0.5–0.8 kg 是主流预算。[BXI5014-19](/zh/motors/bxi5014-19)（0.5 kg、峰值 25 N·m）与 BXI5018-19（0.55 kg、峰值 35 N·m）正落在这个区间；髋部助力需求更高时用 0.8 kg 的 BXI7010-19。

**2. 中空走线有助于提高可靠性。** 外骨骼线缆沿人体布置，关节处反复弯折容易产生疲劳。BXI 全系 6–10 mm 中空孔径可让电源与传感线从关节轴心穿过，减少外部线缆的大幅弯折和勾挂风险。

## 选型对照表

| 应用 | 峰值扭矩需求 | 关键特性 | 建议 BXI 型号 |
| --- | --- | --- | --- |
| 中型四足（10–30 kg） | 20–50 N·m | 反驱吸收落地冲击、低腿部惯量 | [BXI7010-19](/zh/motors/bxi7010-19) / [BXI5018-19](/zh/motors/bxi5018-19) |
| 大型巡检四足（40 kg+ 带载） | 60–150 N·m | 大扭矩爬坡越障、90 A 级供电预算 | [BXI8515-19](/zh/motors/bxi8515-19) |
| 下肢外骨骼髋/膝 | 20–60 N·m | 力控柔顺、0.5–0.8 kg 轻量、中空走线 | [BXI5018-19](/zh/motors/bxi5018-19) / [BXI7010-19](/zh/motors/bxi7010-19) |
| 上肢/轻量外骨骼 | 10–25 N·m | 极致轻量（0.5 kg）、低惯量 | [BXI5014-19](/zh/motors/bxi5014-19) |

> 扭矩区间为工程经验量级，实际需求随整机质量、助力比例与步态变化。四款型号完整参数见[关节电机产品页](/zh/motors/advanced-motors)。

## 常见问题

**机器狗一般用什么电机？** 主流方案是 QDD 准直驱关节电机：无框力矩电机 + 低减速比行星减速器（如 19.5），兼顾扭矩密度与反驱性。10–50 kg 级机器狗单关节峰值扭矩通常在 20–60 N·m，对应 BXI5018/7010 档位；大型巡检犬承力关节用 150 N·m 级的 BXI8515。

**外骨骼机器人电机要满足什么要求？** 三条硬指标：可反向驱动（人带得动机器）、电流环力控（助力柔顺、安全兜底）、轻量化（单关节 0.5–0.8 kg 级）；中空走线可大幅简化穿戴式布线。

**四足和人形的关节电机通用吗？** 高度通用。两者同为腿足式平台，都依赖 QDD 方案，BXI8515-19 既是 Elf 3 人形机器人的髋膝承力关节，也适配大型四足腿部驱动；差别主要在扭矩档位——人形髋膝普遍需 100 N·m 以上，中型四足 50 N·m 档即可覆盖。

人形机器人逐关节的扭矩测算方法，见[《人形机器人髋/膝关节电机选型指南》](/zh/blog/humanoid-hip-knee-joint-motor-guide)；四款型号的三步选型法，见[《BXI 85/70/50 系列关节电机选型指南》](/zh/blog/joint-motor-selection-guide)；需要样机或选型支持，欢迎[联系我们](/zh/contact)。

---

Source: https://www.bxirobotics.com/zh/blog/elf3-humanoid-robot-specifications

# Elf 3 人形机器人完整规格与应用

Elf 3 是 BXI 半醒科技推出的高性能通用人形机器人,具备开放式 SDK 与即插即用系统架构,面向具身智能研究、教育实训与工业场景验证。整机以自研关节电机、PCIE-CANFD 控制架构与 ROS2 软件栈为底座,支持遥操作与自主两种运行模式。本文整理其完整技术规格、各子系统解读与典型应用,供研发与采购团队快速评估。

## 核心规格一览

| 参数 | 数值 |
| --- | --- |
| 全身自由度 | 31（不含手）：单腿 6 + 单臂 7 + 腰 3 + 头 2 |
| 身高 | 1450 mm |
| 宽 × 厚 | 450 × 280 mm |
| 整机重量 | 约 38 kg |
| 最大速度 | 5 m/s（约 18 km/h） |
| 单臂负载 | 5 kg（水平连续）/ 10 kg（屈肘支撑） |
| 电池容量 | 518 Wh |
| 续航 | 约 1 小时（步行） |
| 关节电机 | 31 台 BXI 中空行星关节电机（5014/5018/7010/8515 系列） |
| 控制频率 | ＞1000 Hz（PCIE-CANFD 架构） |
| 运算平台 | Intel Core i7-1370P / 16 GB / 512 GB |
| 传感器 | RealSense D435i 深度相机、IMU、8 麦克风阵列 |
| 软件 | Ubuntu 22.04 + ROS2 SDK + MuJoCo 仿真 |

## 自由度分布

Elf 3 全身 31 个自由度的分布为:每条腿 6 个、每条手臂 7 个、腰部 3 个、头部 2 个。7 自由度手臂支持更接近人类的灵巧操作轨迹,可覆盖更丰富的抓取与作业姿态;3 自由度腰部提升了全身协调与可达作业空间,使弯腰、转体等动作更自然。

## 关节电机:全身动力来源

整机由 31 台自研中空轴行星减速关节电机驱动,覆盖 85/70/50 三个尺寸系列,峰值扭矩从 25 N·m 到 150 N·m:腿部承力关节用大扭矩型号,手臂与末端用轻量型号以压低惯量。全系配备双绝对值编码器(磁式输入 + 电感式输出)与交叉滚子轴承,直接在输出端测量真实角度,实现高精度闭环与上电免归零。各型号参数与选型方法详见[BXI 85/70/50 系列关节电机选型指南](/zh/blog/joint-motor-selection-guide)。

## 控制与算力

整机采用 PCIE-CANFD 控制架构,实现 >1000 Hz 的高频全身控制,为动态平衡与全身协调提供低时延的指令回路。机载算力为 Intel Core i7-1370P、16 GB 内存与 512 GB 存储,可在本体上运行控制与轻量推理负载。

## 感知与软件

- **传感**：RealSense D435i 深度相机提供 RGB-D 视觉,IMU 用于姿态与平衡估计,8 麦克风阵列支持语音交互与声源定位。
- **软件栈**：基于 Ubuntu 22.04 + ROS2,提供硬件控制 SDK 与 MuJoCo 仿真环境,支持从仿真到实机(sim-to-real)的算法迁移,并兼容 MIT 协议的 CANFD 通信。

## 移动与负载

最大速度 5 m/s(约 18 km/h),单臂负载 5 kg(水平连续)/ 10 kg(屈肘支撑),518 Wh 电池支持约 1 小时步行续航,兼顾移动能力与实际作业负载。

## 典型应用场景

- **具身智能研究**：开放式 ROS2 SDK 与 MuJoCo 仿真环境,支持强化学习与运动控制算法的仿真到实机迁移。
- **教育与实训**：即插即用架构降低上手门槛,适合高校与职业院校的人形机器人课程。
- **遥操作与数据采集**：支持遥操作与自主两种模式,可用于操作数据采集与示教。
- **工业场景验证**：作为成熟整机平台,可用于具身方案的可行性验证与二次开发基线。

## 常见问题

**Elf 3 有多少个自由度?** 全身 31 个自由度(不含手):单腿 6、单臂 7、腰 3、头 2。

**Elf 3 跑多快、能扛多重?** 最大速度 5 m/s(约 18 km/h),单臂负载 5 kg(水平连续)/ 10 kg(屈肘支撑)。

**续航多久?** 518 Wh 电池,步行约 1 小时。

**用什么关节电机?** 31 台 BXI 自研中空轴行星关节电机(5014/5018/7010/8515 系列),峰值扭矩 25–150 N·m。

**Elf 3 支持二次开发吗?** 支持。提供基于 ROS2 的硬件控制 SDK 与 MuJoCo 仿真环境,并兼容 MIT 协议的 CANFD 通信。

**能否做 OEM/ODM 定制?** 可以,以 Elf 3 为成熟基线提供品牌化定制,详见[人形机器人 OEM/ODM 全流程](/zh/blog/humanoid-robot-oem-process)。

需要完整规格书、报价或 OEM/ODM 方案,欢迎[联系我们](/zh/contact)。

---

Source: https://www.bxirobotics.com/zh/blog/joint-motor-selection-guide

# BXI 85/70/50 系列关节电机选型指南

BXI 半醒科技的关节电机采用中空轴行星减速结构,全系配备双绝对值编码器(磁式输入 + 电感式输出)与交叉滚子轴承,兼容 MIT 协议的 CAN/CANFD 通信。这套执行器已在多款人形机器人整机上按明确测试条件完成验证,模块化设计同时适配机械臂与腿部等高自由度具身系统。本文给出四款主力型号的完整参数、参数解读与一套可落地的选型方法,帮助研发与采购团队按扭矩和尺寸需求快速决策。

## 完整参数表

| 参数 | BXI8515-19 | BXI7010-19 | BXI5018-19 | BXI5014-19 |
| --- | --- | --- | --- | --- |
| 额定扭矩 | 40 N·m | 15 N·m | 11 N·m | 7 N·m |
| 峰值扭矩 | 150 N·m | 50 N·m | 35 N·m | 25 N·m |
| 额定电压 | 24–48 V | 24–48 V | 24–48 V | 24–48 V |
| 空载转速 | 200 rpm | 200 rpm | 200 rpm | 200 rpm |
| 额定输出转速 | 100 rpm | 100 rpm | 100 rpm | 100 rpm |
| 峰值相电流 | 90 A | 60 A | 30 A | 30 A |
| 减速比 | 19.5 | 19.5 | 19.5 | 19.5 |
| 重量 | 1.4 kg | 0.8 kg | 0.55 kg | 0.5 kg |
| 安装外径 | 99 mm | 81 mm | 64 mm | 64 mm |
| 高度 | 73 mm | 68 mm | 70.5 mm | 66.5 mm |
| 中空孔径 | 10 mm | 9 mm | 6 mm | 6 mm |
| 编码器 | 双绝对值(磁式输入 + 电感式输出) | 同左 | 同左 | 同左 |
| 轴承 | 交叉滚子轴承 | 同左 | 同左 | 同左 |
| 通信接口 | CAN / CANFD | CAN / CANFD | CAN / CANFD | CAN / CANFD |

> 以上为理论值,实际值可能因工况有所偏差。型号前两位对应机座尺寸系列(85/70/50),后缀 **-19** 对应约 19.5 的减速比。

## 关键参数怎么看

选型前先弄清几个参数的含义,能避免"扭矩够不够"的反复试错:

- **额定扭矩 vs 峰值扭矩**：额定扭矩是可长期连续输出的扭矩,用于评估关节的持续保持与稳态负载;峰值扭矩是短时可达的上限,用于覆盖起步、冲击与动态摆动。选型时应让**连续工况落在额定扭矩内**,把**瞬态峰值留给峰值扭矩**。
- **减速比 19.5**：全系采用 19.5 的行星减速比,用转速换扭矩——输出转速约为电机转速的 1/19.5,因此额定输出转速统一为 100 rpm。减速比相同的好处是扭矩主要随机座尺寸递增,选型只需对齐"扭矩档位"。
- **额定电压 24–48 V**：兼容常见机器人母线电压,可在 24 V 与 48 V 系统间复用。
- **峰值相电流**：反映驱动器需要提供的电流能力,机座越大需求越高(BXI8515-19 达 90 A),据此匹配电源与驱动余量。
- **重量与末端惯量**：越靠近末端(腕/手)的关节越应选轻型号,降低末端惯量有利于动态响应与能耗;5014/5018 低至 0.5 kg 正是为此设计。

## 选型三步法

1. **定扭矩档位**：估算关节在最不利姿态下的持续保持扭矩和峰值冲击扭矩,据此圈定型号——保持扭矩对齐额定扭矩,冲击扭矩对齐峰值扭矩。
2. **校尺寸与重量**：在满足扭矩的前提下,优先选更小外径、更轻的型号,尤其是手臂远端关节,以压低惯量与整机重量。
3. **核走线与接口**：确认中空孔径能容纳线缆与传感走线(6–10 mm 可选),并统一 CAN/CANFD 通信与母线电压。

## 怎么选?按关节部位匹配扭矩

- **髋/膝等大扭矩腿部关节**：选 **BXI8515-19**(峰值 150 N·m),提供足够的支撑与动态扭矩。
- **肩/肘等手臂主关节**：选 **BXI7010-19**(峰值 50 N·m),兼顾扭矩与重量。
- **腕/末端等轻量关节**：选 **BXI5018-19 / BXI5014-19**(峰值 35 / 25 N·m),重量低至 0.5 kg,利于减小末端惯量。

## 核心技术:中空轴 + 双绝对值编码器 + 交叉滚子轴承

- **中空轴行星减速**：大直径中空输出为线缆、液压与传感模块腾出走线空间,显著简化整机布线,并增大关节活动范围;行星减速兼顾扭矩密度与结构紧凑。
- **双绝对值编码器**：输入端与输出端各一颗绝对值编码器,可直接在输出端测量真实角度,带来更高的闭环控制精度,并实现上电免归零——开机即知关节位置,无需回零标定。
- **交叉滚子轴承**：单一轴承同时承受径向、轴向与倾覆载荷,提升关节刚度与回转精度,适合承力关节。

## 通信与控制:CAN/CANFD + MIT 协议

全系采用 CAN / CANFD 通信并兼容 MIT 协议,支持力矩 / 速度 / 位置的混合控制,便于在统一总线上实现高带宽、低时延的多关节协同。在高自由度整机上,可搭配 [PCIE-CAN 控制模块](/zh/motors/control-modules)(最多 24 路 CAN、1 KHz 控制)集中调度;[Elf 3 人形机器人](/zh/blog/elf3-humanoid-robot-specifications)即以 PCIE-CANFD 架构实现 >1000 Hz 的整机控制频率。

## 整机验证

85/70/50 系列已在多款人形机器人整机上按明确测试条件完成验证，并应用于 Elf 3 关节系统，覆盖腿部承力与手臂操作等场景。

## 常见问题

**这些电机的扭矩范围是多少?** 全系额定扭矩 7–40 N·m,峰值扭矩 25–150 N·m,覆盖从末端到腿部的关节需求。

**额定扭矩和峰值扭矩怎么用?** 连续保持与稳态负载按额定扭矩选,起步、冲击与动态摆动按峰值扭矩选,留出余量更稳妥。

**支持什么通信协议?** CAN / CANFD,兼容 MIT 协议,支持力矩 / 速度 / 位置混合控制。

**为什么用双绝对值编码器?** 输入与输出端均直接测量真实角度,提升闭环精度,并支持上电免归零,简化标定流程。

**手臂和腿部分别选哪款?** 腿部承力关节选 BXI8515-19(150 N·m),手臂主关节选 BXI7010-19(50 N·m),腕/末端等轻量关节选 BXI5018-19 / BXI5014-19。

需要选型建议或样机,欢迎[联系我们](/zh/contact),也可查看[关节电机产品页](/zh/motors/advanced-motors)。

---

Source: https://www.bxirobotics.com/zh/blog/humanoid-robot-oem-process

## 人形机器人 OEM 与 ODM：从定制需求到整机交付

BXI 半醒科技提供从需求分析到整机交付的 OEM/ODM 服务，覆盖关节电机、控制模块和整机集成。基于现有平台的项目可将 **3–6 个月**作为早期计划参考；实际周期取决于定制深度、供应链、测试和认证要求，并以双方确认的项目计划为准。

## 机器人 OEM 和 ODM 有什么区别？

**OEM 通常侧重按确定的设计与规格制造和集成；ODM 还包括设计与开发。** 对人形机器人项目，关键是先明确谁负责结构设计、关节选型、控制软件和测试，而不是只选择一个合作名称。

| 比较项 | OEM 合作 | ODM 合作 |
| --- | --- | --- |
| 项目输入 | 已有设计、图纸、BOM 或明确的生产规格 | 应用场景、性能目标与定制需求 |
| 主要工作 | 按约定设计采购、制造、组装与测试 | 方案设计、部件选型、样机开发与整机集成 |
| 适合情况 | 已有机器人设计，需要制造与集成支持 | 需要基于现有平台开发自有品牌机器人 |
| 需书面确认 | 设计变更、质量要求、物料替代和验收标准 | 设计责任、知识产权、软件交付和验收标准 |

两种方式都需要在合同中明确具体责任和交付物。BXI 的定制范围可从 [Elf 3 人形机器人平台](/zh/robots/humanoid-robot)与[机器人 ODM/OEM 服务](/zh/custom-robot-oem)开始评估。

## 人形机器人定制报价需要哪些资料？

报价前，建议准备以下信息；尚未确定的项目可标记为待评估。

- **应用与工况**：科研、遥操作或其他目标场景，室内外环境、工作时长和实验要求。
- **本体指标**：尺寸、重量、自由度、关节负载、运动范围、续航与供电方式。
- **软硬件接口**：传感器、算力平台、ROS2 或其他接口、仿真模型以及已有软件的集成边界。
- **采购计划**：样机数量、预计批量、预算、目标交期与交付地点。
- **验证与资料**：验收测试、认证需求、CAD、BOM、技术文档和培训范围。

定制成本通常涉及研发、物料、加工组装、软件集成和验证。**报价需要对应具体配置与数量**；样机单价、批量单价和一次性研发费用应分别确认，起订量也需按项目沟通。

## 如何评估机器人 ODM/OEM 供应商？

建议核对可交付的参数表、接口文档、样机测试范围和制造分工。关节选型应同时检查额定与峰值扭矩、转速、重量和通信接口；软件部分应确认 SDK、仿真模型和部署支持的交付边界。比较报价时使用同一配置与验收条件，并确认维护、备件及设计变更流程。

## 合作流程

| 阶段 | 周期 | 内容 |
| --- | --- | --- |
| 需求收集与保密协议 | 按沟通安排 | 收集需求，按需签署保密协议 |
| 方案范围讨论 | 信息齐备后评估 | 明确规格、扭矩目标与集成边界 |
| 架构与合作框架 | 范围确认后安排 | 确定技术路线、里程碑与合作边界 |
| 研发与交付 | 基线项目可参考 3–6 个月 | 完成定制开发、集成和约定验证 |

### 各阶段说明

- **需求收集与保密协议**：收集应用、预算和交付目标，并按需要签署保密协议。
- **方案范围讨论**：明确规格、扭矩目标、集成边界和验证要求，形成可估算的项目范围。
- **架构与合作框架**：确定技术路线、执行器与控制选型、关键里程碑和双方责任。
- **研发与交付**：进入定制开发与整机集成，完成约定测试后交付。

## 可定制范围

- **外观**：整机造型与品牌视觉,打造区别于公版的自有形象。
- **结构**：本体结构与机构布局,按场景调整尺寸、关节排布与承力设计。
- **性能参数**：基于 BXI 关节电机、控制模块与整机技术进行性能定制,匹配目标扭矩、自由度与控制需求。

## 为什么选择 BXI 的 OEM/ODM?

- **全栈自研**：关节电机(85/70/50 系列,峰值 25–150 N·m)、PCIE-CANFD 控制模块(>1000 Hz)、整机技术均自研可控,定制深度不受外部供应链限制。详见[关节电机选型指南](/zh/blog/joint-motor-selection-guide)。
- **整机集成经验**：电机已集成到多款人形机器人平台，整机可基于 [Elf 3](/zh/blog/elf3-humanoid-robot-specifications) 的现有架构评估定制范围。
- **里程碑管理**：通过范围确认、设计评审和测试节点管理项目进度，周期以双方确认的计划为准。

## 研发与制造布局

定制项目由上海研发团队负责整机研发，深圳制造团队承担机加工与组装，支持从样机到批量交付的工程衔接。

## 常见问题

**OEM/ODM 交期多久?** 基于现有平台、定制范围明确的项目可将 3–6 个月作为早期计划参考；新机构、认证或复杂供应链需求会延长周期，最终以项目计划为准。

**能定制哪些内容?** 可定制外观、结构与性能参数,并基于 BXI 关节电机、控制模块与整机技术做品牌化集成。

**合作流程是怎样的?** 需求收集与保密协议 → 方案范围确认 → 架构与里程碑评审 → 研发、集成、测试与交付。

**哪些因素影响交期?** 定制结构、软件功能、传感器与算力选配、供应链、测试范围和认证要求都会影响周期。

**有成熟的整机基线吗?** 有。以 Elf 3 人形机器人为成熟基线进行品牌化定制,详见 [Elf 3 完整规格](/zh/blog/elf3-humanoid-robot-specifications)。

启动定制项目,请[联系我们](/zh/contact)或查看 [OEM 定制页](/zh/custom-robot-oem)。

---

Source: https://www.bxirobotics.com/zh/blog/hello-world

# 欢迎来到 BXI Robotics

很高兴宣布我们的全新博客正式上线！在这里，我们将分享以下内容的最新动态：

- **人形机器人**：Elf 3 及未来型号的最新进展。
- **关节电机**：深入解析我们的高扭矩密度执行器技术。
- **行业新闻**：具身智能与机器人制造领域的趋势与洞察。

## 为什么开设博客？

我们希望通过这个平台，与大家分享我们的技术积累和产品故事。

1.  **分享技术**：探讨机器人开发中的技术细节。
2.  **发布动态**：第一时间同步产品更新和公司新闻。
3.  **连接社区**：与开发者和机器人爱好者建立更紧密的联系。

敬请期待更多精彩内容！


## Blog posts (English)

---

Source: https://www.bxirobotics.com/blog/humanoid-hip-knee-joint-motor-guide

# How to Choose Hip and Knee Joint Actuators for a Humanoid Robot

**The short answer: for a full-size humanoid robot in the 50–80 kg class, hip and knee joints typically need 100–150 N·m or more of peak torque** to handle standing up, stair climbing, and jumping, while rated (continuous) torque must cover standing posture holds and walking swing. Leg actuators are the first gate in any humanoid design — undersize them and the robot cannot stand up; oversize them and the legs become too heavy to move dynamically. This guide gives a joint-by-joint torque table, a five-point selection method, and a back-of-envelope stair-climbing calculation.

## Joint-by-joint torque requirements and actuator matches

| Joint | Typical peak torque demand | Recommended model (peak torque) |
| --- | --- | --- |
| Hip pitch | 100–150+ N·m | [BXI8515-19](/motors/bxi8515-19) (150 N·m) |
| Hip roll | 80–150 N·m | [BXI8515-19](/motors/bxi8515-19) (150 N·m) |
| Hip yaw | 30–60 N·m | [BXI7010-19](/motors/bxi7010-19) (50 N·m) |
| Knee | 100–150+ N·m | [BXI8515-19](/motors/bxi8515-19) (150 N·m) |
| Ankle | 30–60 N·m | [BXI7010-19](/motors/bxi7010-19) (50 N·m) |
| Shoulder | 30–50 N·m | [BXI7010-19](/motors/bxi7010-19) (50 N·m) |
| Elbow | 20–50 N·m | [BXI7010-19](/motors/bxi7010-19) (50 N·m) |
| Wrist / neck / light loads | 10–35 N·m | [BXI5018-19](/motors/bxi5018-19) (35 N·m) / [BXI5014-19](/motors/bxi5014-19) (25 N·m) |

> These ranges are engineering rules of thumb for 50–80 kg robots; actual demand scales with robot mass, link lengths, and gait. Full specifications for all four models are on the [joint motor product page](/motors/advanced-motors).

## Five things that actually decide the selection

**1. Rated torque for continuous loads, peak torque for transients.** Standing holds and steady walking swing must fit within rated torque (40 N·m on the BXI8515-19); stand-up pushes, stair climbing, and landing impacts are what the 150 N·m peak is for. The most expensive mistake is sizing by peak demand as if it were continuous: an actuator rated for 150 N·m continuous would sit roughly two frame sizes up — about twice the weight and several times the cost — and would wreck leg inertia. The opposite mistake, ignoring the rating, means thermal derating during long standing tasks.

**2. Torque density (N·m/kg) sets your leg weight budget.** Leg actuators are part of the load they carry. The BXI8515-19 delivers 150 N·m peak from 1.4 kg — roughly 107 N·m/kg peak torque density. A lighter actuator at the same torque directly reduces leg swing inertia and whole-robot energy consumption.

**3. Gear ratio affects backdrivability.** [Quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) actuators use a relatively low-ratio [planetary gearbox](/glossary/planetary-gearbox) — 19.5 across the BXI lineup — to reduce backdrive resistance and support output-torque estimation from a calibrated motor model and phase current. High-ratio harmonic drives generally have more backdrive resistance and are often less suitable for highly dynamic, compliant leg joints.

**4. Dual absolute encoders and hollow-shaft cabling.** [Dual absolute encoders](/glossary/dual-absolute-encoder) (magnetic on the input, inductive on the output) measure true output angle directly, so joints power on with no homing routine — re-zeroing 30-plus joints at every boot is a non-starter on a full humanoid. The 10 mm hollow bore routes power and sensor cables through the joint axis, so hip cabling survives large repeated swings instead of fatiguing externally.

**5. Current and voltage budget.** The BXI8515-19 draws up to 90 A peak phase current on a 24–48 V bus. On a 48 V system, every hip and knee driver and its power path must be budgeted for 90 A-class transients — and the bus current spike when all eight large leg joints push off simultaneously is a hard constraint on battery and power distribution design.

## How much torque does a robot knee need? A stair-climbing estimate

A quick order-of-magnitude check: a 70 kg humanoid climbing stairs on one supporting leg, with the center of mass roughly 0.2 m horizontally ahead of the knee (more in a deep squat), needs a knee torque of about:

> τ ≈ m × g × d = 70 kg × 9.8 m/s² × 0.2 m ≈ **137 N·m**

That is a quasi-static approximation — add acceleration terms and the transient demand climbs further, which is exactly why hip and knee peak torque is budgeted at the 150 N·m level. Continuous torque for walking swing and standing holds usually lands in the 20–40 N·m range, within the BXI8515-19's 40 N·m rating.

## FAQ

**How much torque does a humanoid robot knee need?** For a 50–80 kg robot, plan for 100–150+ N·m peak (stairs, sit-to-stand) and 20–40 N·m continuous. The BXI8515-19 (40 N·m rated / 150 N·m peak) is sized for exactly this duty cycle.

**Harmonic drive or planetary gearbox for hip joints?** For legs, choose a low-ratio planetary (QDD) design: it is backdrivable, tolerates impacts, and enables force-controlled compliant landing. 100:1 harmonic drives resist backdriving and their flexsplines dislike shock loads — better suited to robot arms.

**Why do humanoid legs use QDD actuators?** A low gear ratio improves backdrivability and allows output torque to be estimated from phase current after calibrating motor torque constant, gearbox efficiency, and friction. Some joint-control tasks can work without a dedicated joint torque sensor, while high-accuracy contact measurement and safety-critical applications may still require foot or six-axis force/torque sensors.

For the full four-model spec table and a three-step selection workflow, see the [BXI 85/70/50 Joint Motor Selection Guide](/blog/joint-motor-selection-guide), or [contact us](/contact) for samples and sizing support.

---

Source: https://www.bxirobotics.com/blog/quadruped-exoskeleton-actuator-guide

# Robot Dog Motors and Exoskeleton Actuators: A Selection Guide

**Bottom line: quadrupeds and exoskeletons both benefit from high [torque density](/glossary/torque-density), backdrivability, and compliant [force control](/glossary/force-control), making low-ratio [quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) actuators a common option.** The final architecture still depends on load, speed, control accuracy, and safety requirements.

## What motor does a robot dog use?

A typical quadruped carries **12 active joints**: two at each hip (abduction/adduction plus flexion/extension) and one at each knee. Modern quadrupeds use both hydraulic and electric architectures. Low-ratio electric drives are common on small and mid-size platforms that prioritize maintenance, efficiency, and controllability, but no single architecture covers every quadruped.

A defining load case is **repeated ground impact**. Backdrivability can reduce peak loads in the drivetrain and lets the controller use calibrated current-based torque estimation for compliant response. High-ratio harmonic drives generally have more reflected inertia and friction, so they are often less suitable for highly dynamic legs. BXI joint motors use a 19.5 planetary reduction to balance torque multiplication and backdrivability.

**Torque sizing for quadrupeds.** For robot dogs in the 10–50 kg class, per-joint peak torque demand typically lands in the **20–60 N·m** range, with the knee usually the most demanding joint due to its longer moment arm. Larger inspection-class quadrupeds carrying payloads push well beyond that:

- **Mid-size quadrupeds (10–30 kg)**: the [BXI7010-19](/motors/bxi7010-19) (50 N·m peak at 0.8 kg) suits hips and knees, while the [BXI5018-19](/motors/bxi5018-19) (35 N·m peak at 0.55 kg) covers lighter axes such as hip abduction. Keeping leg mass low matters twice here — leg inertia sets both stride frequency and battery life.
- **Large inspection quadrupeds (40 kg+ with payload)**: the [BXI8515-19](/motors/bxi8515-19) (150 N·m peak, 40 N·m rated) handles slope climbing, obstacle negotiation, and payload carriage. Its 90 A peak phase current on a 24–48 V bus is also the number to budget the 12-joint power system around.

## What must an exoskeleton actuator deliver?

An exoskeleton works in direct contact with a **person**. Excessive backdrive resistance increases wearer effort, so backdrivability and compliant control are important design requirements. Motor-current and model-based torque estimation can support the control loop, but safety should also include mechanical limits, fault handling, and appropriate force or torque sensing where required.

Three product classes dominate: **rehabilitation exoskeletons** (clinical, trajectory-led with force-control safety limits), **industrial assist exoskeletons** (load-handling relief, roughly 20–60 N·m of peak hip/knee assist), and **consumer hiking exoskeletons** — rental units at mountain tourist sites have become a widely reported trend, and this class pushes weight and cost to the floor.

Two selection criteria are specific to wearables:

**1. Actuator mass is a hard constraint.** Every gram hangs on the wearer. The sweet spot for a wearable joint is **0.5–0.8 kg per actuator** — exactly where the [BXI5014-19](/motors/bxi5014-19) (0.5 kg, 25 N·m peak) and BXI5018-19 (0.55 kg, 35 N·m peak) sit. When hip assist demands more torque, the 0.8 kg BXI7010-19 stays inside a wearable weight budget while tripling available peak torque over the 5014.

**2. Hollow-shaft routing can improve cable reliability.** Exoskeleton cables flex repeatedly at joints. BXI's 6–10 mm hollow bore routes power and sensor lines through the rotation axis, reducing large external bends and snagging risks.

## Selection table: quadruped and exoskeleton applications

| Application | Peak torque range | Key requirements | Suggested BXI model |
| --- | --- | --- | --- |
| Mid-size quadruped (10–30 kg) | 20–50 N·m | Backdrivable impact absorption, low leg inertia | [BXI7010-19](/motors/bxi7010-19) / [BXI5018-19](/motors/bxi5018-19) |
| Large inspection quadruped (40 kg+, payload) | 60–150 N·m | High torque for slopes and obstacles, 90 A power budget | [BXI8515-19](/motors/bxi8515-19) |
| Lower-limb exoskeleton hip/knee | 20–60 N·m | Compliant force control, 0.5–0.8 kg mass, hollow-bore cabling | [BXI5018-19](/motors/bxi5018-19) / [BXI7010-19](/motors/bxi7010-19) |
| Upper-limb / lightweight exoskeleton | 10–25 N·m | Minimum mass (0.5 kg), low inertia | [BXI5014-19](/motors/bxi5014-19) |

> Torque ranges are engineering orders of magnitude; actual demand varies with machine mass, assist ratio, and gait. Full specifications for all four models are on the [joint motor product page](/motors/advanced-motors).

## FAQ

**What motor does a robot dog typically use?** A QDD (quasi-direct-drive) joint actuator — a frameless torque motor with a low planetary reduction such as 19.5, balancing torque density against backdrivability. Per-joint peak torque for 10–50 kg robot dogs is typically 20–60 N·m (BXI5018/7010 class); large inspection quadrupeds use 150 N·m-class actuators like the BXI8515-19.

**What are the requirements for an exoskeleton motor?** Three hard ones: backdrivability (the wearer must be able to move the machine), current-loop force control (compliant assist with a safety floor), and low mass (0.5–0.8 kg per joint). A hollow bore for through-axis cable routing greatly simplifies wearable integration.

**Are quadruped and humanoid joint motors interchangeable?** Largely yes. Both are legged platforms built on QDD actuators — the BXI8515-19 serves as the load-bearing hip/knee joint of the Elf 3 humanoid and equally suits large quadruped legs. The difference is torque tier: humanoid hips and knees generally need 100 N·m+, while a mid-size quadruped is covered at the 50 N·m tier.

For per-joint torque estimation on humanoids, see the [humanoid hip/knee joint motor guide](/blog/humanoid-hip-knee-joint-motor-guide); for the three-step method across all four models, see the [BXI 85/70/50 series selection guide](/blog/joint-motor-selection-guide). Need samples or sizing support? [Contact us](/contact).

---

Source: https://www.bxirobotics.com/blog/elf3-humanoid-robot-specifications

# Elf 3 Humanoid Robot: Full Specifications and Applications

The Elf 3 is BXI Robotics' high-performance, general-purpose humanoid robot, featuring an open SDK and a ready-to-deploy system architecture for embodied-AI research, education, and industrial validation. It is built on in-house joint motors, a PCIE-CANFD control architecture, and a ROS2 software stack, and supports both teleoperation and autonomous modes. This article summarizes its full specifications, a breakdown of each subsystem, and typical use cases for engineering and procurement teams.

## Specifications at a glance

| Parameter | Value |
| --- | --- |
| Total degrees of freedom | 31 (excl. hands): 6 per leg + 7 per arm + 3 waist + 2 head |
| Height | 1450 mm |
| Width × Depth | 450 × 280 mm |
| Weight | ≈ 38 kg |
| Max speed | 5 m/s (~18 km/h) |
| Single-arm payload | 5 kg (horizontal, continuous) / 10 kg (elbow-supported) |
| Battery | 518 Wh |
| Runtime | ≈ 1 hour (walking) |
| Joint motors | 31 × BXI hollow-shaft planetary motors (5014/5018/7010/8515 series) |
| Control rate | > 1000 Hz (PCIE-CANFD architecture) |
| Onboard compute | Intel Core i7-1370P / 16 GB / 512 GB |
| Sensors | RealSense D435i depth camera, IMU, 8-mic array |
| Software | Ubuntu 22.04 + ROS2 SDK + MuJoCo simulation |

## Degrees of freedom

The 31 DOF break down as 6 per leg, 7 per arm, 3 in the waist, and 2 in the head. The 7-DOF arms enable more human-like manipulation trajectories and cover a wider range of grasping and working poses, while the 3-DOF waist improves whole-body coordination and reachable workspace, making motions like bending and twisting more natural.

## Joint motors: the power behind every joint

The robot is driven by 31 in-house hollow-shaft planetary joint motors across three size series (85/70/50), with peak torque from 25 N·m to 150 N·m: high-torque models for load-bearing leg joints, lighter models for arms and end joints to keep inertia low. All use dual absolute encoders (magnetic input + inductive output) and cross-roller bearings, measuring the true output angle directly for high-accuracy closed-loop control and power-on without homing. For per-model specs and a selection method, see the [BXI 85/70/50 joint motor selection guide](/blog/joint-motor-selection-guide).

## Control and compute

The robot uses a PCIE-CANFD control architecture to achieve a >1000 Hz whole-body control rate, providing a low-latency command loop for dynamic balance and whole-body coordination. Onboard compute is an Intel Core i7-1370P with 16 GB of memory and 512 GB of storage, capable of running control and lightweight inference workloads on the body.

## Sensing and software

- **Sensing:** the RealSense D435i depth camera provides RGB-D vision, the IMU is used for pose and balance estimation, and the 8-mic array supports voice interaction and sound-source localization.
- **Software stack:** built on Ubuntu 22.04 + ROS2, with a hardware control SDK and a MuJoCo simulation environment that support sim-to-real algorithm transfer, plus MIT-protocol-compatible CANFD communication.

## Mobility and payload

A top speed of 5 m/s (~18 km/h), single-arm payload of 5 kg (horizontal, continuous) / 10 kg (elbow-supported), and a 518 Wh battery for about 1 hour of walking runtime balance mobility with real working payload.

## Typical use cases

- **Embodied-AI research:** an open ROS2 SDK and MuJoCo simulation environment support sim-to-real transfer for reinforcement learning and motion-control algorithms.
- **Education and training:** the ready-to-deploy architecture lowers the barrier to entry for university and vocational humanoid-robotics courses.
- **Teleoperation and data collection:** both teleoperation and autonomous modes are supported for manipulation data collection and demonstration.
- **Industrial validation:** as a mature whole-robot platform, it serves as a feasibility-validation and secondary-development baseline for embodied solutions.

## FAQ

**How many degrees of freedom does the Elf 3 have?** 31 DOF excluding hands: 6 per leg, 7 per arm, 3 waist, 2 head.

**How fast is it and how much can it carry?** Top speed of 5 m/s (~18 km/h), single-arm payload 5 kg (horizontal, continuous) / 10 kg (elbow-supported).

**What is the runtime?** A 518 Wh battery for about 1 hour of walking.

**Which joint motors does it use?** 31 in-house BXI hollow-shaft planetary joint motors (5014/5018/7010/8515 series), with 25–150 N·m peak torque.

**Can the Elf 3 be developed on?** Yes. It ships with a ROS2-based hardware control SDK and a MuJoCo simulation environment, and uses MIT-protocol-compatible CANFD communication.

**Is OEM/ODM customization available?** Yes — branded customization is offered with the Elf 3 as a mature baseline. See the [humanoid robot OEM/ODM process](/blog/humanoid-robot-oem-process).

For the full datasheet, a quote, or an OEM/ODM solution, [contact us](/contact).

---

Source: https://www.bxirobotics.com/blog/joint-motor-selection-guide

# BXI 85/70/50 Joint Motor Selection Guide

BXI Robotics joint motors use a hollow-shaft planetary reduction design, with dual absolute encoders (magnetic input + inductive output) and cross-roller bearings across the lineup, plus MIT-protocol-compatible CAN/CANFD communication. The actuators have been validated on multiple humanoid platforms under defined test conditions, and their modular design fits both robotic arms and legs in high-DOF embodied systems. This guide gives the full specs of the four main models, a breakdown of what each parameter means, and a practical workflow so engineering and procurement teams can select by torque and size.

## Full spec table

| Parameter | BXI8515-19 | BXI7010-19 | BXI5018-19 | BXI5014-19 |
| --- | --- | --- | --- | --- |
| Rated torque | 40 N·m | 15 N·m | 11 N·m | 7 N·m |
| Peak torque | 150 N·m | 50 N·m | 35 N·m | 25 N·m |
| Rated voltage | 24–48 V | 24–48 V | 24–48 V | 24–48 V |
| No-load speed | 200 rpm | 200 rpm | 200 rpm | 200 rpm |
| Rated output speed | 100 rpm | 100 rpm | 100 rpm | 100 rpm |
| Peak phase current | 90 A | 60 A | 30 A | 30 A |
| Reduction ratio | 19.5 | 19.5 | 19.5 | 19.5 |
| Weight | 1.4 kg | 0.8 kg | 0.55 kg | 0.5 kg |
| Installation diameter | 99 mm | 81 mm | 64 mm | 64 mm |
| Height | 73 mm | 68 mm | 70.5 mm | 66.5 mm |
| Hollow bore | 10 mm | 9 mm | 6 mm | 6 mm |
| Encoder | Dual absolute (magnetic input + inductive output) | Same | Same | Same |
| Bearing | Cross-roller bearing | Same | Same | Same |
| Communication | CAN / CANFD | CAN / CANFD | CAN / CANFD | CAN / CANFD |

> Values are theoretical; actual values may vary by operating conditions. The first two digits denote the frame-size series (85/70/50), and the **-19** suffix corresponds to the ~19.5 reduction ratio.

## How to read the key parameters

Understanding a few parameters up front avoids trial-and-error over "is the torque enough":

- **Rated vs peak torque:** rated torque is what the motor can output continuously — use it for sustained holding and steady-state loads; peak torque is the short-duration ceiling — use it for start-up, impact, and dynamic swings. Keep **continuous operation within rated torque** and **reserve peak torque for transients**.
- **19.5 reduction ratio:** the whole lineup uses a 19.5 planetary ratio, trading speed for torque — output speed is about 1/19.5 of motor speed, which is why rated output speed is a uniform 100 rpm. A common ratio means torque scales mainly with frame size, so selection comes down to matching a "torque tier".
- **24–48 V rated voltage:** compatible with common robot bus voltages and reusable across 24 V and 48 V systems.
- **Peak phase current:** reflects the current the drive must supply — larger frames demand more (90 A on the BXI8515-19) — size your power supply and drive headroom accordingly.
- **Weight and end-effector inertia:** joints closer to the end (wrist/hand) should use lighter models; lower end inertia helps dynamic response and energy use — the 5014/5018, as light as 0.5 kg, are built for exactly this.

## A three-step selection workflow

1. **Set the torque tier:** estimate the continuous holding torque and peak impact torque in the worst-case pose, then shortlist models — match holding torque to rated torque and impact torque to peak torque.
2. **Check size and weight:** among models that meet the torque requirement, prefer the smaller-diameter, lighter one — especially for distal arm joints — to cut inertia and overall weight.
3. **Verify routing and interface:** confirm the hollow bore can carry your cables and sensor wiring (6–10 mm options), and standardize CAN/CANFD communication and bus voltage.

## How to choose: match torque to joint location

- **High-torque leg joints (hip/knee):** choose the **BXI8515-19** (150 N·m peak) for sufficient support and dynamic torque.
- **Primary arm joints (shoulder/elbow):** choose the **BXI7010-19** (50 N·m peak) to balance torque and weight.
- **Lightweight distal joints (wrist/end):** choose the **BXI5018-19 / BXI5014-19** (35 / 25 N·m peak), as light as 0.5 kg to reduce end-effector inertia.

## Core technology: hollow shaft + dual absolute encoders + cross-roller bearings

- **Hollow-shaft planetary reduction:** the large-diameter hollow output frees routing space for cables, hydraulics, and sensor modules, greatly simplifying whole-robot wiring and increasing joint range of motion; planetary reduction balances torque density with a compact structure.
- **Dual absolute encoders:** one absolute encoder on each of the input and output sides measures the true output angle directly, delivering higher closed-loop control accuracy and enabling power-on without homing — the joint position is known at boot, with no zeroing routine.
- **Cross-roller bearings:** a single bearing handles radial, axial, and moment loads simultaneously, improving joint stiffness and rotational accuracy — ideal for load-bearing joints.

## Communication and control: CAN/CANFD + MIT protocol

The whole lineup uses CAN / CANFD and is MIT-protocol compatible, supporting hybrid torque / speed / position control for high-bandwidth, low-latency multi-joint coordination on a single bus. On high-DOF robots, pair them with the [PCIE-CAN control modules](/motors/control-modules) (up to 24 CAN buses, 1 kHz control) for centralized scheduling; the [Elf 3 humanoid robot](/blog/elf3-humanoid-robot-specifications) uses a PCIE-CANFD architecture to reach a >1000 Hz whole-robot control rate.

## Validated on whole robots

The 85/70/50 series has been validated on multiple humanoid platforms under defined test conditions and is used across the Elf 3 joint system, from load-bearing legs to arm manipulation.

## FAQ

**What is the torque range?** Rated torque 7–40 N·m and peak torque 25–150 N·m across the lineup, covering joints from end-effector to leg.

**How do I use rated vs peak torque?** Select by rated torque for continuous holding and steady-state loads, and by peak torque for start-up, impact, and dynamic swings — leaving headroom for safety.

**Which protocols are supported?** CAN / CANFD, MIT-protocol compatible, with hybrid torque / speed / position control.

**Why dual absolute encoders?** Both input and output sides measure the true angle directly, improving closed-loop accuracy and enabling power-on without homing, which simplifies calibration.

**Which model for arms vs legs?** Use the BXI8515-19 (150 N·m) for load-bearing leg joints, the BXI7010-19 (50 N·m) for primary arm joints, and the BXI5018-19 / BXI5014-19 for lightweight wrist/end joints.

For selection advice or samples, [contact us](/contact) or see the [joint motors page](/motors/advanced-motors).

---

Source: https://www.bxirobotics.com/blog/humanoid-robot-oem-process

## Humanoid Robot OEM vs ODM: From Requirements to Delivery

BXI Robotics provides OEM/ODM services from requirements analysis through whole-robot delivery, covering joint actuators, control modules, and system integration. Projects based on an existing platform may use **3–6 months** as an early planning reference; actual schedules depend on customization, supply chain, testing, and certification requirements and are confirmed during project scoping.

## What is the difference between robot OEM and ODM?

**OEM typically focuses on manufacturing and integration to an agreed design; ODM also includes design and development.** For humanoid robots, clarify who owns mechanical design, actuator selection, control software and testing before choosing an engagement label.

| Consideration | OEM engagement | ODM engagement |
| --- | --- | --- |
| Project inputs | Existing design, drawings, BOM or production specifications | Application, performance targets and customization requirements |
| Main work | Procurement, manufacturing, assembly and testing to the agreed design | Design, component selection, prototype development and integration |
| Typical fit | An existing robot design needs manufacturing and integration support | A branded robot needs development on an existing platform |
| Agree in writing | Design changes, quality requirements, material substitutions and acceptance tests | Design responsibility, intellectual property, software deliverables and acceptance tests |

Both approaches require explicit responsibilities and deliverables. Start scoping with the [Elf 3 humanoid platform](/robots/humanoid-robot) and [BXI robot OEM/ODM services](/custom-robot-oem).

## What is needed for a custom humanoid robot quote?

Prepare the following inputs, marking unresolved items for assessment:

- **Application and operating conditions:** research, teleoperation or another intended use, environment, duty cycle and experiment requirements.
- **Robot requirements:** dimensions, weight, degrees of freedom, joint loads, motion range, runtime and power supply.
- **Hardware and software interfaces:** sensors, compute, ROS2 or other interfaces, simulation models and integration boundaries with existing software.
- **Procurement plan:** prototype quantity, expected production volume, budget, target date and delivery location.
- **Validation and documentation:** acceptance tests, certification needs, CAD, BOM, technical documentation and training scope.

Cost can include development, materials, machining and assembly, software integration and validation. **A quote must correspond to a configuration and quantity.** Confirm prototype pricing, production pricing and one-time development fees separately; minimum quantities are project-specific.

## How should a robot OEM/ODM supplier be evaluated?

Review deliverable specifications, interface documentation, prototype tests and manufacturing responsibilities. Actuator selection should account for rated and peak torque, speed, weight and communication interfaces. Confirm what the SDK, simulation models and deployment support include. Compare quotes against the same configuration and acceptance criteria, including maintenance, spare parts and design-change procedures.

## Engagement process

| Stage | Timeline | What happens |
| --- | --- | --- |
| Intake & NDA | Scheduled with the project team | Collect requirements and sign an NDA if needed |
| Application scoping | Assessed after inputs are complete | Define specs, torque targets, and integration scope |
| Architecture and framework | Scheduled after scope review | Agree on architecture, milestones, and responsibilities |
| Development & delivery | Existing-platform projects may reference 3–6 months | Complete development, integration, and agreed validation |

### Stage details

- **Intake & NDA:** collect application, budget, and delivery goals and sign an NDA if required.
- **Application scoping:** define specifications, torque targets, integration boundaries, and validation requirements.
- **Architecture and framework:** agree on the technical route, actuator and control selection, milestones, and responsibilities.
- **Development & delivery:** complete custom development, integration, and the agreed tests before delivery.

## What can be customized

- **Exterior:** overall industrial design and brand visuals, for an identity distinct from the reference design.
- **Structure:** body structure and mechanical layout, with dimensions, joint arrangement, and load-bearing design tuned to the application.
- **Performance:** performance parameters tuned on BXI joint motors, control modules, and whole-robot technology to match target torque, DOF, and control needs.

## Why choose BXI for OEM/ODM?

- **Full-stack in-house:** joint motors (85/70/50 series, 25–150 N·m peak), PCIE-CANFD control modules (>1000 Hz), and whole-robot technology are all developed and controlled in-house, so customization depth is not limited by an external supply chain. See the [joint motor selection guide](/blog/joint-motor-selection-guide).
- **System-integration experience:** the actuators have been integrated into multiple humanoid platforms, and the [Elf 3](/blog/elf3-humanoid-robot-specifications) architecture can serve as a reference when scoping customization.
- **Milestone management:** scope reviews, design reviews, and validation gates make schedules measurable; the committed timeline is defined in the project plan.

## R&D and manufacturing footprint

Custom projects are engineered by the Shanghai R&D team, with machining and assembly supported by the Shenzhen manufacturing team, connecting prototype development with batch delivery.

## FAQ

**What is the OEM/ODM lead time?** Existing-platform projects with a defined scope may use 3–6 months as an early planning reference. New mechanisms, certification, or complex supply-chain requirements can extend the schedule.

**What can be customized?** Exterior, structure, and performance parameters, with branded integration built on BXI's joint motors, control modules, and whole-robot technology.

**What does the process look like?** Intake and NDA → scope confirmation → architecture and milestone review → development, integration, testing, and delivery.

**What affects the lead time?** Mechanical customization, software scope, sensor and compute options, supply chain, validation depth, and certification requirements all affect the schedule.

**Is there a mature whole-robot baseline?** Yes. Branded customization is built on the Elf 3 humanoid robot as a mature baseline — see the [Elf 3 full specifications](/blog/elf3-humanoid-robot-specifications).

To start a custom project, [contact us](/contact) or see the [OEM customization page](/custom-robot-oem).

---

Source: https://www.bxirobotics.com/blog/hello-world

# Welcome to BXI Robotics

We are thrilled to announce the official launch of our new blog! Here, we will share the latest updates on:

- **Humanoid Robots**: The latest progress on Elf 3 and future models.
- **Joint Motors**: In-depth analysis of our high-torque density actuator technology.
- **Industry News**: Trends and insights in embodied AI and robot manufacturing.

## Why a Blog?

We hope to use this platform to share our technology stack and product stories with you.

1.  **Sharing Technology**: Discussing technical details in robot development.
2.  **Releasing Updates**: Synchronizing product updates and company news immediately.
3.  **Connecting Community**: Building closer connections with developers and robotics enthusiasts.

Stay tuned for more exciting content!


## Glossary (Chinese)

---

Source: https://www.bxirobotics.com/zh/glossary/coreless-motor

# 空心杯电机是什么？

**空心杯电机**（Coreless Motor，又称无铁芯电机）的转子是一个**无铁芯的杯状自支撑绕组**，导线绕制成薄壁圆筒直接在磁场中旋转。去掉铁芯后没有铁损、没有齿槽转矩，转子惯量比同尺寸有铁芯电机低一个量级，电气时间常数通常在 **1 ms 以下**，启停和换向响应极快。

## 为什么灵巧手偏爱空心杯

[灵巧手](/zh/glossary/dexterous-hand)的手指关节空间只有十几毫米直径，却要做高频抓握动作：

- **体积小**：直径 4–16 mm 的空心杯配微型行星/丝杠即可驱动一根手指；
- **低惯量高响应**：手指开合频率可达数 Hz 甚至更高，低惯量转子跟得上；
- **无齿槽顿感**：抓取力平滑，利于指尖力控制；
- **效率高**：常见 70%–90%，对手掌内散热受限的场景很关键。

代价是杯状绕组散热差、过载能力弱，持续扭矩通常只有 mN·m 量级，必须配高减速比才能输出可用的抓握力。

## 空心杯 vs 无框力矩电机：机器人上的分工

| 维度 | 空心杯电机 | [无框力矩电机](/zh/glossary/frameless-torque-motor) |
| --- | --- | --- |
| 转子结构 | 无铁芯杯状绕组 | 有铁芯定转子分离套件 |
| 典型直径 | 4–40 mm | 40–120 mm |
| 扭矩量级 | mN·m 级（需高减速比） | N·m 级（小减速比即可） |
| 典型位置 | 灵巧手手指、微型执行器 | 臂/腿关节电机 |

一句话分工：**手指用空心杯，臂腿关节用无框力矩电机**。BXI 的 [85/70/50 系列关节电机](/zh/motors/advanced-motors)即基于无框电机 + 行星减速的[准直驱](/zh/glossary/quasi-direct-drive)方案，负责躯干与四肢的大扭矩输出，与末端灵巧手的空心杯驱动互补。

---

Source: https://www.bxirobotics.com/zh/glossary/crossed-roller-bearing

# 交叉滚子轴承是什么？

**交叉滚子轴承**是一种把圆柱滚子以相邻两颗互相垂直（90° 交叉）的方式排列在 V 形滚道中的精密轴承。由于滚子交替朝向两个方向，单个轴承就能同时承受径向载荷、双向轴向载荷与倾覆力矩。

## 为什么机器人关节偏爱它？

普通深沟球轴承主要承受径向载荷，要覆盖复合载荷通常需要成对布置、加大结构。而机器人关节的载荷天然是复合的——以人形机器人髋关节为例，支撑体重（轴向）、摆腿离心力（径向）与整腿悬伸产生的弯矩（倾覆）同时作用。交叉滚子轴承用**一个轴承位解决全部三种载荷**：

- **高刚度**：滚子线接触优于滚珠点接触，抗形变能力强；
- **高回转精度**：适合作为[双绝对值编码器](/zh/glossary/dual-absolute-encoder)输出端的测量基准；
- **节省空间**：单轴承方案让[关节电机](/zh/glossary/joint-motor)更薄、更轻。

BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)全系在输出端配置交叉滚子轴承，覆盖从腿部承力到手臂灵巧操作的载荷需求。

---

Source: https://www.bxirobotics.com/zh/glossary/degrees-of-freedom

# 自由度是什么？

**自由度**（Degrees of Freedom，DoF）指机器人可独立控制的运动轴数量。每个可主动驱动的旋转或平移关节记为一个自由度，通常对应一台[关节电机](/zh/glossary/joint-motor)。

## 自由度决定动作能力

- **6 自由度**是机械臂在三维空间达到"任意位置 + 任意姿态"的最低要求（3 个定位置 + 3 个定姿态）；
- **7 自由度**手臂引入一个冗余自由度，可在末端位姿不变的情况下调整肘部位置，绕开障碍、优化姿态——这正是人类手臂的构型；
- **人形机器人**需要全身协调，自由度通常在 30 个以上。

## 实例：Elf 3 的 31 个自由度

BXI [Elf 3 人形机器人](/zh/robots/humanoid-robot)全身 31 个自由度（不含手）：单腿 6、单臂 7、腰 3、头 2。腿部 6 自由度覆盖行走所需的髋（3）、膝（1）、踝（2）；7 自由度手臂提供类人操作冗余；3 自由度腰部扩大可达作业空间。

自由度越多，对执行器数量、控制总线带宽（见 [MIT 协议](/zh/glossary/mit-protocol-can)）与整机重量的压力越大，因此自由度布局是人形机器人设计的核心取舍之一。

---

Source: https://www.bxirobotics.com/zh/glossary/dexterous-hand

# 灵巧手是什么？

**灵巧手**（Dexterous Hand）是模仿人手结构与功能的多自由度机器人末端执行器：通过多个手指与关节的协同运动，实现抓取、捏取、在手内调整物体姿态等精细操作。相比只有开/合两种状态的二指夹爪，灵巧手能适应形状各异的物体和工具，是[人形机器人](/zh/robots/humanoid-robot)走向通用操作的关键部件。

## 自由度：欠驱动 vs 全驱动

人手约有 20 余个[自由度](/zh/glossary/degrees-of-freedom)，灵巧手的设计在"接近人手"与"工程可靠"之间权衡：

- **欠驱动方案（约 6-12 DOF 主动驱动）**：电机数量少于关节数量，多个关节由同一驱动源联动，靠机械结构自适应贴合物体。成本低、可靠性高，能覆盖大部分抓取任务；
- **全驱动方案（20+ DOF）**：每个关节独立可控，能完成手内操作（in-hand manipulation）等高难度动作，但结构复杂、成本和维护门槛显著更高，多见于科研平台。

## 驱动方案

- **空心杯电机直驱/微型减速**:利用空心杯电机的高功率密度，把驱动器塞进手指或手掌，响应快、结构紧凑；
- **腱绳传动**：电机置于前臂，通过腱绳（类似人类肌腱）远程拉动手指，手部轻量化，但存在摩擦与弹性带来的控制误差；
- **连杆传动**：刚性连杆传递运动，精度和刚度好，自由度布置相对受限。

## 与人形机器人的关系

灵巧手决定了人形机器人"能干什么活"。在[具身智能](/zh/glossary/embodied-ai)研究中，灵巧操作数据主要靠[遥操作](/zh/glossary/teleoperation)（数据手套、外骨骼、VR）采集，供模仿学习策略训练。手部自由度越高，可学习的任务空间越大，但数据采集与控制难度也同步上升——这也是当前多数落地场景仍从夹爪或欠驱动灵巧手起步的原因。

---

Source: https://www.bxirobotics.com/zh/glossary/dual-absolute-encoder

# 双绝对值编码器是什么？

**双绝对值编码器**是高端[关节电机](/zh/glossary/joint-motor)的编码器配置：在电机转子（输入端、减速前）与关节输出法兰（输出端、减速后）各布置一颗绝对值编码器，两端同时直接测量角度。

## 为什么需要两颗？

只在电机端装编码器时，输出角度要靠"电机角度 ÷ 减速比"推算，而减速器的背隙、弹性形变与装配误差都会让推算值偏离真实关节角度。输出端编码器则**直接测量减速后的真实角度**，闭环控制建立在真值之上：

- **控制精度更高**：消除背隙与形变造成的角度误差；
- **上电免归零**：绝对值编码器断电后仍保留绝对位置，开机即知关节姿态，无需回零标定流程；
- **双端冗余**：输入/输出角度互相校验，可检测减速器打滑等异常。

## 实现方式

BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)全系采用磁式编码器测输入端、电感式编码器测输出端的组合，两种感测原理互不干扰，在紧凑结构内实现真双编码闭环。

---

Source: https://www.bxirobotics.com/zh/glossary/embodied-ai

# 具身智能是什么？

**具身智能**（Embodied AI）是指智能体拥有物理身体，通过身体与真实环境的交互来完成感知、决策与行动的人工智能范式。与只处理文本、图像的"离身"模型不同，具身智能必须在物理世界中闭环：看到 → 理解 → 动手 → 观察结果 → 修正。

## 为什么人形机器人是主要载体

人类环境（楼梯、门把手、工具、工位）都是为人类身体设计的。人形形态让机器人**无需改造环境**即可复用这些设施，也让海量的人类动作数据（视频、动作捕捉、[遥操作](/zh/glossary/teleoperation)示教）可以直接迁移为训练数据。

## 技术栈构成

- **本体**：高动态硬件是前提——高扭矩密度[关节电机](/zh/glossary/joint-motor)、高频控制总线与全身[自由度](/zh/glossary/degrees-of-freedom)布局；
- **感知**：视觉、深度、IMU、触觉等多模态传感；
- **决策**：大模型/强化学习策略，常经 [Sim-to-Real](/zh/glossary/sim-to-real) 从仿真迁移到实机；
- **数据**：遥操作采集真机数据，用于模仿学习。

BXI 的 [Elf 3 人形机器人](/zh/robots/humanoid-robot)提供 ROS2 SDK 与 MuJoCo 仿真环境，正是面向具身智能研究的开放平台。

---

Source: https://www.bxirobotics.com/zh/glossary/force-control

# 机器人力控是什么？

**力控**（Force Control）指机器人关节直接控制**输出力/力矩**而不是只控制位置，让机器人在接触环境和人时表现出可控的"柔软度"。走路落地、双臂搬运、人机协作都离不开力控——纯位置控制的关节遇到意外接触会硬顶到底，力控关节则会顺势让步。

## 三种实现方式

- **力矩控制**：直接给定关节力矩目标，最底层也最快，[MIT 协议](/zh/glossary/mit-protocol-can)的 `T_ff` 前馈力矩即此用途；
- **阻抗控制**：给定"虚拟弹簧-阻尼"参数（刚度 Kp、阻尼 Kd），位置偏差自动转成力矩输出，是腿式机器人最常用的方式，MIT 协议一帧指令即可同时下发 Kp/Kd 与力矩前馈；
- **导纳控制**：测量外力、输出位置修正，常用于末端装力传感器的机械臂，带宽受位置环限制。

## 要不要力矩传感器？

高减速比关节的摩擦、效率变化和迟滞会降低电机电流到输出扭矩估算的准确度，因此高精度应用常配置关节或末端力传感器。[准直驱（QDD）](/zh/glossary/quasi-direct-drive)关节减速比较低，在完成电机转矩常数、减速器效率和摩擦标定后，可结合相电流估算输出扭矩。该方法适合部分本体感知与阻抗控制任务，但不能替代所有安全关键或高精度接触测量。BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)通过 19.5 减速比[行星方案](/zh/glossary/planetary-gearbox)、MIT 协议 CAN 指令与[双绝对值编码器](/zh/glossary/dual-absolute-encoder)支持关节级控制。

---

Source: https://www.bxirobotics.com/zh/glossary/frameless-torque-motor

# 无框力矩电机是什么？

**无框力矩电机**（Frameless Torque Motor）是只交付**定子和转子两个部件**的电机：没有自带外壳、输出轴和轴承，由用户把定转子直接压装进自己的关节结构件里。它省去了"电机壳套关节壳"的冗余结构，让关节更短、更轻、散热路径更直接，是[一体化关节模组](/zh/glossary/joint-motor)与人形机器人执行器的主流电机形态。

## 为什么关节模组用无框电机

- **结构融合**：定子直接热装进关节壳体，转子直接连[行星减速器](/zh/glossary/planetary-gearbox)太阳轮，传动链最短，无框设计可让关节体积缩小 30% 左右；
- **大直径薄饼形**：力矩电机走多极对数、大气隙半径路线，低转速高扭矩，天然匹配关节"低速大扭矩"的需求，也是高[扭矩密度](/zh/glossary/torque-density)的来源；
- **中空同轴**：定转子环形结构便于实现[中空走线](/zh/glossary/hollow-shaft-motor)，线缆与传感线从关节中心穿过。

## 内转子与外转子

无框力矩电机分内转子（转子在内圈，散热好、惯量低）与外转子（转子在外圈，同体积扭矩更大、更扁平）两种拓扑。高动态腿式机器人常用外转子方案换取更高扭矩密度；BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)即基于无框力矩电机 + 19.5 行星减速的[准直驱](/zh/glossary/quasi-direct-drive)架构，配[双绝对值编码器](/zh/glossary/dual-absolute-encoder)构成完整一体化关节。

---

Source: https://www.bxirobotics.com/zh/glossary/harmonic-drive

# 谐波减速器是什么？

**谐波减速器**（Harmonic Drive，学名应变波齿轮 Strain Wave Gear）通过椭圆波发生器让薄壁**柔轮**产生弹性变形、与刚轮差齿啮合来传动，单级即可实现 **50–160 的大减速比**，背隙接近零、体积重量小，是精密传动三大件（谐波、RV、行星）之一。

## 谐波减速器和行星减速器的区别

| 维度 | 谐波减速器 | [行星减速器](/zh/glossary/planetary-gearbox) |
| --- | --- | --- |
| 单级减速比 | 50–160 | 3–10（多级可叠加） |
| 背隙 | 近零 | 小但存在 |
| 抗冲击 | 柔轮怕冲击过载 | 齿轮实心，抗冲击强 |
| 反向可驱动 | 困难（摩擦大） | 小减速比时良好 |
| 典型位置 | 机械臂腕/肘、人形手臂 | 腿部、高动态关节 |

一句话结论：**要精度选谐波，要抗冲击和力控选行星**。

## 机器人怎么选

工业与协作机械臂的定位精度需求让谐波成为手臂关节标配；但人形与四足机器人的腿部要承受落地冲击、还要靠电流反推外力做力控，高减速比谐波的摩擦与柔轮疲劳成为短板，所以腿部普遍走[准直驱（QDD）](/zh/glossary/quasi-direct-drive)行星路线——BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)即采用 19.5 减速比行星方案，兼顾[扭矩密度](/zh/glossary/torque-density)与反驱性。基座级大扭矩场景则常用 RV 减速器（摆线针轮），构成"基座 RV、手臂谐波、腿部行星"的常见分工。

---

Source: https://www.bxirobotics.com/zh/glossary/hollow-shaft-motor

# 中空轴电机是什么？

**中空轴电机**是指输出轴中心带有贯通孔的电机/执行器结构。与实心轴方案相比，它允许线缆、液压管路与传感线束直接从关节旋转中心穿过，而不必绕关节外侧走线。

## 为什么机器人需要中空轴？

高自由度机器人（如 31 自由度的人形机器人）关节数量多、串联层级深，如果每个关节的线缆都从外侧绕行：

- 线束会在关节反复弯折处疲劳损坏；
- 外露线缆限制关节活动范围，还可能与环境剐蹭；
- 整机布线复杂度和装配工时急剧上升。

从中心孔走线则一次性解决这些问题——这是[关节电机](/zh/glossary/joint-motor)向一体化、高集成演进的关键设计。

## 工程要点

中空孔径越大，可通过的线束越多，但电机内径、编码器与减速器的布置难度也随之上升。BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)提供 6–10 mm 中空孔径，配合[行星减速器](/zh/glossary/planetary-gearbox)在扭矩密度与走线空间之间取得平衡。

---

Source: https://www.bxirobotics.com/zh/glossary/joint-motor

# 关节电机是什么？

**关节电机**（也称机器人关节模组、一体化执行器）是将无框力矩电机、减速器、编码器、驱动器与轴承集成在同一紧凑结构内的机器人动力部件。它直接安装在机器人的关节处，输出受控的扭矩与转角，是人形机器人、四足机器人与机械臂的"肌肉"。

## 核心组成

- **无框力矩电机**：提供原始扭矩，直接决定动力上限。
- **减速器**：以转速换扭矩，常见方案有[行星减速器](/zh/glossary/planetary-gearbox)与谐波减速器。
- **编码器**：测量转角用于闭环控制，高端方案采用[双绝对值编码器](/zh/glossary/dual-absolute-encoder)。
- **轴承**：承受负载并保证回转精度，承力关节多用[交叉滚子轴承](/zh/glossary/crossed-roller-bearing)。
- **驱动器**：执行电流环/速度环/位置环控制，通过 CAN/CANFD 等总线接收指令（见 [MIT 协议](/zh/glossary/mit-protocol-can)）。

## 关键指标

选型时最重要的参数是**额定扭矩**（可持续输出）与**峰值扭矩**（短时上限）、重量、外形尺寸、减速比与通信接口。一台人形机器人通常需要 20–40 台不同扭矩档位的关节电机——例如 BXI Elf 3 全身使用 31 台 [85/70/50 系列关节电机](/zh/motors/advanced-motors)，覆盖 25–150 N·m 峰值扭矩。

选型方法可参考[关节电机选型指南](/zh/blog/joint-motor-selection-guide)。

---

Source: https://www.bxirobotics.com/zh/glossary/mit-protocol-can

# MIT 协议是什么？

**MIT 协议**（MIT Mode）是源自 MIT Mini Cheetah 开源腿式机器人项目的 CAN 总线电机控制协议。它的核心是把一条控制指令压缩进一帧 CAN 报文，同时下发五个量：目标位置 p、目标速度 v、位置增益 Kp、速度增益 Kd 与前馈扭矩 τ。

## 力位混合控制

电机端按 `τ_out = Kp·(p - p_actual) + Kd·(v - v_actual) + τ` 计算输出扭矩。通过调节增益，同一条协议可以覆盖三种控制模式：

- **纯力矩控制**：Kp = Kd = 0，只发前馈扭矩——高动态力控的基础；
- **位置控制**：Kp、Kd 正常取值，像伺服一样跟踪位置；
- **阻抗控制**：中间增益让关节表现出"弹簧-阻尼"特性，柔顺接触环境。

## 为什么它成了事实标准

协议开源、指令紧凑（单帧完成）、天然支持力控，被大量[关节电机](/zh/glossary/joint-motor)厂商兼容，形成了跨品牌的通用控制接口。BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)全系兼容 MIT 协议的 CAN/CANFD 通信，配合 [PCIE-CAN 控制模块](/zh/motors/control-modules)可实现 >1000 Hz 的整机高频控制。

---

Source: https://www.bxirobotics.com/zh/glossary/planetary-gearbox

# 行星减速器是什么？

**行星减速器**是一种由太阳轮（输入）、若干行星轮与内齿圈组成的齿轮传动机构：行星轮同时绕自身轴线自转并围绕太阳轮公转，像行星绕恒星运动，因此得名。它把电机的高转速转换为低转速、大扭矩输出。

## 结构优势

- **多齿同时啮合**：负载由多个行星轮分担，扭矩密度高、承载强。
- **输入输出同轴**：结构紧凑，天然适合做成圆柱形的[关节电机](/zh/glossary/joint-motor)。
- **传动效率高**：单级效率通常可达 95% 以上，高于谐波减速器。
- **易做中空结构**：中心可开贯通孔实现[中空轴走线](/zh/glossary/hollow-shaft-motor)。

## 与谐波减速器的取舍

谐波减速器减速比大（50–160）、零背隙，但效率较低、刚度有限、价格高；行星减速器减速比较小（单级约 3–10，多级组合可达约 20），有少量背隙，但效率、抗冲击性与成本占优。对高动态的人形机器人腿部关节，行星方案的抗冲击与效率优势尤其重要。

BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)全系采用 19.5 减速比的行星方案，额定输出转速统一为 100 rpm，选型时只需对齐扭矩档位。

---

Source: https://www.bxirobotics.com/zh/glossary/quasi-direct-drive

# 准直驱执行器是什么？

**准直驱执行器**（Quasi-Direct Drive，QDD）是一类采用大直径、大扭矩电机搭配**小减速比**（通常 5–20）减速器的机器人执行器方案，最早由 MIT Cheetah 等腿式机器人项目推广。"准直驱"意为接近直接驱动（无减速）但保留少量减速。

## 核心特性：反向可驱动

减速比越大，从输出端反向拖动电机就越困难（摩擦与惯量按减速比平方放大）。小减速比让关节保持**反向可驱动性**（backdrivability）：

- 外部冲击能被电机"让"掉而不是硬抗，落地、碰撞时保护齿轮；
- 完成转矩常数、效率和摩擦标定后，可结合电机电流估计关节输出扭矩，降低部分场景对专用关节力矩传感器的依赖；
- 力控带宽高，适合跳跃、奔跑等高动态运动。

## 与高减速比方案的取舍

高减速比（谐波、RV）方案通常具有较高输出扭矩和保持能力，但反驱阻力也更大；QDD 更偏向动态响应和柔顺控制。许多动态腿式机器人采用 QDD。BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)使用 19.5 减速比[行星方案](/zh/glossary/planetary-gearbox)，支持 MIT 协议控制（见 [MIT 协议](/zh/glossary/mit-protocol-can)）。

---

Source: https://www.bxirobotics.com/zh/glossary/reflected-inertia

# 反射惯量是什么？

**反射惯量**（Reflected Inertia，又称折算惯量）是把电机转子惯量折算到减速器输出端后的等效惯量：**J_reflected = J_rotor × N²**，其中 N 是减速比。惯量按减速比的**平方**放大——减速比 100 的关节，输出端"看到"的转子惯量是原来的 **10000 倍**；减速比 10 则只有 100 倍。

## 为什么高减速比关节"手感硬"

外界推动关节输出端时，必须连带加速被平方放大的转子惯量：

- **抗冲击差**：落地或碰撞瞬间，冲击能量来不及被控制器响应，全部砸在减速器齿面和柔轮上，这是高减速比关节在腿部易损坏的主因之一；
- **反驱阻力大**：反射惯量叠加摩擦，用手推几乎推不动，物理柔顺性差；
- **力控带宽受限**：控制器要"演"出柔软，必须先克服巨大的等效惯量，主动柔顺的带宽和保真度都被拉低。

一句话可引用：**减速比翻倍，反射惯量翻四倍——这就是腿部关节不敢用大减速比的物理原因。**

## QDD 怎么压低反射惯量

[准直驱（QDD）](/zh/glossary/quasi-direct-drive)路线的核心取舍就在这里：用大直径高[扭矩密度](/zh/glossary/torque-density)电机换取小减速比（典型 6–20），把反射惯量比谐波方案（50–160）压低 **1–2 个数量级**，让关节天然可反驱、冲击由电机侧惯量"软"着陆，电流环即可估计外力。BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)选择 19.5 减速比的行星方案，正是在输出扭矩与反射惯量之间取的工程平衡点。

---

Source: https://www.bxirobotics.com/zh/glossary/robot-data-collection

# 具身智能数据采集是什么？

**具身智能数据采集**（数采，Robot Data Collection）是指在机器人执行任务的过程中，同步记录"观测（摄像头图像、关节状态）+ 动作（关节指令）"配对数据的过程，通常由人类通过[遥操作](/zh/glossary/teleoperation)示教完成。这些数据是训练 [VLA 模型](/zh/glossary/vla-model)与模仿学习策略的原料——模型能学到什么，取决于采到了什么。

## 主从臂遥操作采集流程

主从臂（leader-follower）是常见的真机数采方案：

1. 操作员拖动轻量化的**主臂**（leader），**从臂**（follower）实时复现主臂的关节运动；
2. 系统以固定频率（通常 30-50 Hz 以上）同步记录多路摄像头图像、从臂关节状态与动作指令；
3. 每完成一次任务保存为一条轨迹（episode），一个任务通常采集数十到数百条；
4. 数据经筛选、对齐后进入训练管线。

当主从臂在关节空间一一对应时，示教较直观，也可减少在线运动学映射的复杂度。BXI UpperBody 1 [双臂遥操作平台](/zh/robots/robotic-arms)采用双主臂 + 双从臂架构，可接入 ROS2 数据录制工作流；具体接口与数据格式以项目配置为准。

## 真机数据 vs 仿真数据

- **真机数据**：包含真实的摩擦、形变、光照与传感噪声，训练出的策略可直接部署，但采集成本高、依赖人力；
- **仿真数据**：可大规模并行生成、成本低，但与实机存在差距，需通过 [Sim-to-Real](/zh/glossary/sim-to-real) 技术弥合。

实践中两者常混合使用：仿真提供规模，真机数据锚定真实分布。

## 模仿学习如何消费这些数据

以 ALOHA/ACT 为代表的路线，直接用主从臂采集的双臂轨迹训练策略：ACT（Action Chunking with Transformers）把未来一小段动作作为整体预测。所需示教数量取决于任务难度、数据质量和泛化范围；这类方法说明“遥操作数采 + 模仿学习”是[具身智能](/zh/glossary/embodied-ai)策略训练的一条实用途径。

---

Source: https://www.bxirobotics.com/zh/glossary/rv-reducer

# RV 减速器是什么？

**RV 减速器**（RV Reducer，属摆线针轮 Cycloidal 家族）采用**行星轮系前级 + 摆线针轮后级的两级复合结构**，摆线轮与针齿多点同时啮合，单机减速比 **30–200**，扭转刚性和抗冲击过载能力（常见 2–5 倍额定）在精密减速器三大件中最强，额定扭矩覆盖数十至数千 N·m。名称来自纳博特斯克的 RV 系列，现已成为该结构的通称。

## 谐波、RV、行星：三分工

| 维度 | RV 减速器 | [谐波减速器](/zh/glossary/harmonic-drive) | [行星减速器](/zh/glossary/planetary-gearbox) |
| --- | --- | --- | --- |
| 单机减速比 | 30–200 | 50–160 | 3–10（可多级） |
| 刚性/抗冲击 | 高 | 柔轮对冲击较敏感 | 良好 |
| 背隙 | 约 1 arcmin | 近零 | 小但存在 |
| 重量体积 | 大而重 | 小巧 | 中等 |
| 反向可驱动 | 困难 | 困难 | 小减速比时良好 |
| 典型位置 | 工业机器人基座/大臂 | 腕/小臂、人形手臂 | 人形/四足腿部 |

常见应用倾向是：**工业机器人基座使用 RV，轻量手臂使用谐波，强调反驱性的腿部关节使用低减速比行星方案**；最终选择仍取决于载荷、精度、重量和冲击工况。

## 为什么人形机器人腿部较少使用 RV

六轴工业机器人固定在地面，基座关节要承受整臂负载和悬臂力矩，RV 的高刚性和承载能力很适合这类工况。人形机器人则必须同时携带关节本体：在同类扭矩等级下，RV 方案通常更重，会降低整机[扭矩密度](/zh/glossary/torque-density)；较高摩擦和较差反驱性也会增加基于电机电流估算输出扭矩的不确定性。因此，强调轻量、反驱与柔顺控制的腿部关节通常更倾向小减速比行星[准直驱（QDD）](/zh/glossary/quasi-direct-drive)方案。BXI [85/70/50 系列关节电机](/zh/motors/advanced-motors)采用 19.5 减速比行星结构，峰值扭矩最高 150 N·m。

---

Source: https://www.bxirobotics.com/zh/glossary/sim-to-real

# Sim-to-Real 是什么？

**Sim-to-Real**（仿真到实机）是指先在物理仿真器（如 MuJoCo、Isaac Gym/Lab）中训练机器人控制策略，再把训练好的策略部署到真实机器人上的技术路线。真机训练又慢又贵还会摔坏硬件，而仿真可以并行跑成千上万个机器人、以超实时速度积累"经验"，这让强化学习训练人形机器人行走等技能成为可能。

## 核心挑战：现实差距

仿真永远不完全等于现实（摩擦、延迟、电机特性、传感噪声的差异），这个差异称为 **Sim-to-Real Gap（现实差距）**。主流缓解手段：

- **域随机化**：训练时随机扰动仿真参数（质量、摩擦、延迟），逼策略学会鲁棒性；
- **精确系统辨识**：把真机的电机响应、惯量参数测准后写回仿真；
- **执行器建模**：对[关节电机](/zh/glossary/joint-motor)的扭矩-速度特性单独建模，这是腿式机器人迁移成败的关键之一。

## 对硬件的要求

策略在实机上以数百 Hz 频率输出关节指令，要求执行器具备高带宽力控（见[准直驱](/zh/glossary/quasi-direct-drive)与 [MIT 协议](/zh/glossary/mit-protocol-can)）与低时延总线。BXI [Elf 3 人形机器人](/zh/robots/humanoid-robot)内置 MuJoCo 仿真环境与 ROS2 SDK，配合 >1000 Hz 的 PCIE-CANFD 控制架构，支持从仿真到实机的完整工作流。

---

Source: https://www.bxirobotics.com/zh/glossary/six-axis-force-torque-sensor

# 六维力传感器是什么？

**六维力传感器**（Six-Axis Force/Torque Sensor）在一个测量点上同时输出**三个正交方向的力（Fx、Fy、Fz）和三个方向的力矩（Mx、My、Mz）**，共六个分量，常见方案基于应变片或电容原理。量程、精度和采样率应按具体型号数据表确认。

## 装在哪里、干什么用

- **腕部**：机械臂末端与工具之间，用于精密装配、打磨抛光、拖动示教——凡是需要感知末端接触力的[力控](/zh/glossary/force-control)任务；
- **踝部**：人形机器人脚踝，测量地面反作用力与压力中心（ZMP/CoP），是行走平衡控制的关键输入；
- **测试台**：关节电机与整机的扭矩标定、碰撞测试。

## 有了电流环力控，还需要外置力传感器吗？

[准直驱（QDD）](/zh/glossary/quasi-direct-drive)关节可在完成电机和传动标定后，结合相电流估算输出扭矩。该方法适合部分关节级柔顺控制，但精度会受摩擦、温度、效率和模型误差影响。两类方案的大致分工是：

| 场景 | 电流环力估计（QDD） | 外置六维力传感器 |
| --- | --- | --- |
| 腿部落地缓冲、全身柔顺 | 可用于控制反馈，精度取决于标定 | 按控制与安全需求配置 |
| 末端精密装配 | 通常无法直接测量末端六维受力 | 通常更合适 |
| 脚底 ZMP/地面反力测量 | 只能间接估计 | 更直接可靠 |
| 硬件与集成 | 不增加专用关节传感器 | 增加传感器、标定与布线 |

一句话结论：**QDD 电流估算适合关节级柔顺反馈，六维力传感器适合直接测量末端或足底受力**；两者可以按精度和安全需求组合使用。

---

Source: https://www.bxirobotics.com/zh/glossary/teleoperation

# 遥操作是什么？

**遥操作**（Teleoperation）指人类操作员通过操控设备实时远程控制机器人运动、由机器人在现场执行任务的技术。操作端（主端）捕捉人的动作意图，机器人端（从端）复现动作并回传视觉等反馈，形成人在回路的控制闭环。

## 两重价值

1. **直接干活**：在自主算法尚未成熟的任务上，遥操作让机器人立即具备实用能力（危险环境作业、远程装配、演示）。
2. **采集数据**：遥操作产生的"观察-动作"配对轨迹是**模仿学习最高质量的训练数据**。当前[具身智能](/zh/glossary/embodied-ai)模型的主流训练管线，正是靠大规模遥操作示教数据驱动的。

## 常见形态

- **同构主从臂**：主端是与从端结构相同/相似的小型臂，映射直接、精度高；
- **VR/动捕**：头显与手柄或全身动捕驱动人形机器人全身动作；
- **半自主遥操作**：人给高层指令，底层平衡与轨迹由机器人自主完成。

BXI 的 [UpperBody 1 双臂平台](/zh/robots/robotic-arms)配套即插即用操作控制台，[Elf 3 人形机器人](/zh/robots/humanoid-robot)同时支持遥操作与自主两种模式，可直接用于操作数据采集。

---

Source: https://www.bxirobotics.com/zh/glossary/torque-density

# 扭矩密度是什么？

**扭矩密度**（Torque Density）指执行器单位重量能输出的扭矩，单位 **N·m/kg**，分峰值扭矩密度与额定扭矩密度两种口径。它是机器人[关节电机](/zh/glossary/joint-motor)最核心的指标：关节电机装在机器人肢体上，自身重量就是别的关节要驮的负载——扭矩密度每提高一分，整机就能更轻、更快、续航更久。

## 怎么计算与比较

扭矩密度 = 输出扭矩 ÷ 执行器总重（含减速器与编码器）。比较时务必对齐口径：

- **峰值 vs 额定**：厂商宣传多用峰值口径，数值可达额定口径的 3 倍以上；
- **是否含驱动器**：一体化关节模组含驱动板，裸电机数字不可直接对比；
- 当前一体化关节的峰值扭矩密度普遍在 30–70 N·m/kg，顶尖产品超过 100 N·m/kg。

以 BXI [85/70/50 系列](/zh/motors/advanced-motors)为例（峰值口径）：[BXI8515-19](/zh/motors/bxi8515-19) 为 150 N·m ÷ 1.4 kg ≈ **107 N·m/kg**，[BXI7010-19](/zh/motors/bxi7010-19) 为 50 N·m ÷ 0.8 kg ≈ 63 N·m/kg，[BXI5018-19](/zh/motors/bxi5018-19) 约 64 N·m/kg，[BXI5014-19](/zh/motors/bxi5014-19) 为 50 N·m/kg。

## 高扭矩密度怎么来的

大直径、多极对数的外转子力矩电机提供基础扭矩，[行星减速器](/zh/glossary/planetary-gearbox)以较小的重量代价放大扭矩；[准直驱（QDD）](/zh/glossary/quasi-direct-drive)路线用中小减速比在扭矩密度与反驱性之间取得平衡。谐波方案扭矩密度口径可以更高，但牺牲抗冲击与反驱能力，见[谐波减速器](/zh/glossary/harmonic-drive)。

---

Source: https://www.bxirobotics.com/zh/glossary/vla-model

# VLA 模型是什么？

**VLA 模型**（Vision-Language-Action Model，视觉-语言-动作模型）是一类端到端的具身智能大模型：输入是摄像头图像与自然语言指令（如"把杯子放进抽屉"），输出直接是机器人动作序列（关节角、末端位姿或夹爪指令）。它把感知、语义理解与运动生成压缩进同一个网络，是[具身智能](/zh/glossary/embodied-ai)领域最受关注的策略架构之一。

## 与 LLM 的关系

VLA 通常在视觉-语言模型（VLM）的基础上扩展而来：先复用大语言模型的语义与常识能力，再把输出头改造为动作 token 或连续动作，从而让机器人"听得懂人话、看得懂场景"。可以把 VLA 理解为 LLM 家族向物理世界的延伸——语言模型预测下一个词，VLA 预测下一个动作。

## 为什么需要真机数据

互联网上有海量文本和图像，却几乎没有"图像 + 指令 → 关节动作"的配对数据。因此 VLA 训练高度依赖真机示教数据，主要通过[遥操作](/zh/glossary/teleoperation)采集；仿真数据可以补充规模，但需经 [Sim-to-Real](/zh/glossary/sim-to-real) 弥合与实机的差距。数据的多样性（任务、场景、本体）往往比模型参数量更能决定泛化能力。

## 代表工作

- **RT-2**（Google DeepMind）：将 VLM 与机器人动作 token 联合训练，验证了"网络知识迁移到操作"的可行性；
- **OpenVLA**：开源 VLA 模型，基于大规模开放机器人数据集训练，便于社区微调；
- **π0**（Physical Intelligence）：采用流匹配生成连续动作，面向跨本体的通用操作策略。

在真实平台上部署与微调 VLA，需要可稳定执行高频动作指令的本体——BXI 的[人形机器人](/zh/robots/humanoid-robot)与[双臂机器人平台](/zh/robots/robotic-arms)均提供 ROS2 接口，适合作为 VLA 研究的实机载体。

---

Source: https://www.bxirobotics.com/zh/glossary/world-model

# 世界模型是什么？

**世界模型**（World Model）在机器人语境下，是指智能体内部学习到的环境动力学预测模型：给定当前观测（图像、状态）与将要执行的动作，预测环境接下来会变成什么样。有了它，机器人可以在"脑内"推演动作后果——先想象、再行动——而不必每一步都在真实世界中试错。

## 与 VLA、强化学习的关系

- **对比 [VLA 模型](/zh/glossary/vla-model)**：VLA 直接从观测和指令映射到动作，是"反应式"策略；世界模型则显式建模"动作 → 后果"，可用于规划和推演，两者可以互补——世界模型为 VLA 提供想象与验证能力。
- **对比强化学习**：在模型基强化学习（model-based RL）中，策略大量在世界模型内部"做梦"训练，只需少量真实交互，大幅降低真机采样成本；这与在仿真器中训练再做 [Sim-to-Real](/zh/glossary/sim-to-real) 迁移的思路一脉相承——世界模型可以看作一个从数据中学出来的、可微分的"神经仿真器"。

## 为什么成为 2026 年具身智能热点

一是视频生成模型的进展表明，大模型能够从海量视频中学到相当程度的物理规律，为"通用世界模型"提供了技术底座；二是真机数据昂贵稀缺，业界希望用世界模型放大有限的[遥操作](/zh/glossary/teleoperation)数据——在学到的模型里生成、评估更多虚拟经验；三是[具身智能](/zh/glossary/embodied-ai)对安全性要求高，先在内部模型中预演再上真机，是降低试错风险的自然路径。

无论哪条技术路线，世界模型最终都要落到真实本体上验证。BXI 的[人形机器人](/zh/robots/humanoid-robot)提供 MuJoCo 仿真与 ROS2 SDK，支持"仿真/模型内训练 → 实机验证"的完整研究闭环。


## Glossary (English)

---

Source: https://www.bxirobotics.com/glossary/coreless-motor

# What Is a Coreless Motor?

A **coreless motor** (ironless motor) uses a **self-supporting cup-shaped winding without an iron core** as its rotor — the copper cylinder itself spins in the magnetic field. Removing the iron eliminates iron losses and cogging torque, cuts rotor inertia by roughly an order of magnitude versus an iron-core motor of the same size, and pushes the electrical time constant **below 1 ms** for very fast start-stop response.

## Why Dexterous Hands Use Coreless Motors

A [dexterous hand](/glossary/dexterous-hand) finger has only 10–20 mm of diameter to work with, yet must grasp at high frequency:

- **Compact**: a 4–16 mm coreless motor plus a micro planetary gearbox or leadscrew drives one finger;
- **Low inertia, fast response**: finger open/close cycles of several Hz are achievable;
- **No cogging**: smooth grasp force, better fingertip force control;
- **Efficient**: typically 70%–90%, important with the limited heat dissipation inside a palm.

The trade-off: the cup winding dissipates heat poorly and tolerates little overload — continuous torque sits in the mN·m range, so a high gear ratio is mandatory for usable grip force.

## Coreless vs. Frameless Torque Motor: Division of Labor

| Dimension | Coreless motor | [Frameless torque motor](/glossary/frameless-torque-motor) |
| --- | --- | --- |
| Rotor | Ironless cup winding | Iron-core stator/rotor kit |
| Typical diameter | 4–40 mm | 40–120 mm |
| Torque scale | mN·m (needs high ratio) | N·m (low ratio suffices) |
| Typical location | Fingers, micro actuators | Arm/leg joint actuators |

One-line split: **coreless for fingers, frameless torque motors for arm and leg joints**. BXI's [85/70/50-series joint actuators](/motors/advanced-motors) take the frameless-motor-plus-planetary [quasi-direct-drive](/glossary/quasi-direct-drive) route for high-torque limbs, complementing coreless-driven hands at the end effector.

---

Source: https://www.bxirobotics.com/glossary/crossed-roller-bearing

# What Is a Crossed Roller Bearing?

A **crossed roller bearing** is a precision bearing whose cylindrical rollers alternate orientation at 90° to each other inside a V-shaped raceway. Because adjacent rollers face opposite directions, one bearing simultaneously carries radial loads, bidirectional axial loads, and overturning (moment) loads.

## Why Robot Joints Prefer It

Standard deep-groove ball bearings mainly take radial load; covering combined loads usually requires paired bearings and a bigger structure. Robot joint loads are inherently combined — a humanoid hip joint simultaneously supports body weight (axial), leg-swing centrifugal force (radial), and the bending moment of a cantilevered leg (overturning). A crossed roller bearing handles **all three load types in a single bearing position**:

- **High stiffness**: roller line contact resists deformation better than ball point contact.
- **High rotational precision**: a stable reference for output-side measurement by [dual absolute encoders](/glossary/dual-absolute-encoder).
- **Space savings**: the single-bearing design keeps [joint motors](/glossary/joint-motor) thinner and lighter.

The BXI [85/70/50-series joint motors](/motors/advanced-motors) fit crossed roller bearings at the output across the whole lineup, from load-bearing legs to dexterous arms.

---

Source: https://www.bxirobotics.com/glossary/degrees-of-freedom

# What Are Degrees of Freedom?

**Degrees of freedom (DoF)** count the independently controllable motion axes of a robot. Each actively driven rotary or prismatic joint counts as one DoF, typically corresponding to one [joint motor](/glossary/joint-motor).

## DoF Determines Capability

- **6 DoF** is the minimum for a robot arm to reach "any position + any orientation" in 3D space (3 for position + 3 for orientation);
- **7 DoF** adds a redundant axis, letting the elbow move while the hand stays fixed — dodging obstacles and optimizing posture, exactly the configuration of the human arm;
- **Humanoid robots** need whole-body coordination and typically exceed 30 DoF.

## Example: The 31 DoF of Elf 3

The BXI [Elf 3 humanoid robot](/robots/humanoid-robot) has 31 DoF excluding hands: 6 per leg, 7 per arm, 3 in the waist, 2 in the head. The 6-DoF legs cover walking's hip (3), knee (1), and ankle (2); the 7-DoF arms provide human-like manipulation redundancy; the 3-DoF waist expands the reachable workspace.

More DoF means more actuators, more control-bus bandwidth (see [MIT protocol](/glossary/mit-protocol-can)), and more weight — making DoF layout one of the central trade-offs in humanoid design.

---

Source: https://www.bxirobotics.com/glossary/dexterous-hand

# What Is a Dexterous Hand?

A **dexterous hand** is a multi-degree-of-freedom robotic end-effector modeled on the structure and function of the human hand: coordinated motion of multiple fingers and joints enables grasping, pinching, and in-hand reorientation of objects. Unlike a two-finger gripper with only open/close states, a dexterous hand adapts to objects and tools of arbitrary shape, making it the key component for [humanoid robots](/robots/humanoid-robot) moving toward general-purpose manipulation.

## Degrees of Freedom: Underactuated vs Fully Actuated

The human hand has roughly 20+ [degrees of freedom](/glossary/degrees-of-freedom); dexterous-hand design trades off human-likeness against engineering reliability:

- **Underactuated (about 6-12 actively driven DOF)**: fewer motors than joints, with multiple joints coupled to one drive and mechanical compliance conforming to the object. Lower cost, higher reliability, covers most grasping tasks;
- **Fully actuated (20+ DOF)**: every joint independently controlled, enabling in-hand manipulation and other demanding skills — at the price of complexity, cost, and maintenance, so mostly found on research platforms.

## Actuation Approaches

- **Coreless-motor direct/geared drive**: exploits the power density of coreless motors to fit actuators inside fingers or the palm — fast response, compact structure;
- **Tendon drive**: motors sit in the forearm and pull fingers via tendons (like human anatomy), keeping the hand light, but friction and cable elasticity introduce control error;
- **Linkage drive**: rigid linkages transmit motion with good precision and stiffness, at some cost in DOF layout flexibility.

## Role in Humanoid Robotics

The dexterous hand defines what work a humanoid can actually do. In [embodied AI](/glossary/embodied-ai) research, dexterous manipulation data is mostly collected through [teleoperation](/glossary/teleoperation) — data gloves, exoskeletons, VR — to train imitation-learning policies. More hand DOF means a larger learnable task space, but also harder data collection and control, which is why most deployments today still start from grippers or underactuated hands.

---

Source: https://www.bxirobotics.com/glossary/dual-absolute-encoder

# What Is a Dual Absolute Encoder?

A **dual absolute encoder** configuration is the encoder architecture of high-end [joint motors](/glossary/joint-motor): one absolute encoder on the motor rotor (input side, before reduction) and one on the output flange (output side, after reduction), measuring angle directly at both ends.

## Why Two Encoders?

With only a motor-side encoder, output angle must be inferred as "motor angle ÷ reduction ratio" — and gearbox backlash, elastic deformation, and assembly tolerance all pull that estimate away from the true joint angle. An output-side encoder **measures the real post-reduction angle directly**, so the control loop closes on ground truth:

- **Higher precision**: angle errors from backlash and deformation are eliminated.
- **Zero-calibration startup**: absolute encoders retain position through power loss, so the robot knows every joint pose at power-on with no homing routine.
- **Redundancy**: input and output angles cross-check each other, catching anomalies such as gearbox slip.

## Implementation

The BXI [85/70/50-series joint motors](/motors/advanced-motors) pair a magnetic encoder on the input with an inductive encoder on the output — two sensing principles that don't interfere — achieving true dual-encoder closed loop in a compact package.

---

Source: https://www.bxirobotics.com/glossary/embodied-ai

# What Is Embodied AI?

**Embodied AI** (embodied intelligence) is the paradigm in which an intelligent agent has a physical body and accomplishes perception, decision-making, and action through interaction with the real environment. Unlike "disembodied" models that only process text or images, embodied AI must close the loop in the physical world: see → understand → act → observe the result → correct.

## Why Humanoids Are the Primary Platform

Human environments — stairs, door handles, tools, workstations — are designed for the human body. The humanoid form lets robots reuse this infrastructure **without modifying the environment**, and lets vast human motion data (video, motion capture, [teleoperation](/glossary/teleoperation) demonstrations) transfer directly into training data.

## The Stack

- **Body**: high-dynamic hardware is the prerequisite — torque-dense [joint motors](/glossary/joint-motor), high-frequency control buses, and whole-body [degrees of freedom](/glossary/degrees-of-freedom);
- **Perception**: multimodal sensing — vision, depth, IMU, touch;
- **Decision**: foundation-model or reinforcement-learning policies, often transferred from simulation via [Sim-to-Real](/glossary/sim-to-real);
- **Data**: teleoperated real-robot data collection for imitation learning.

The BXI [Elf 3 humanoid robot](/robots/humanoid-robot) ships with a ROS2 SDK and MuJoCo simulation — an open platform built for embodied-AI research.

---

Source: https://www.bxirobotics.com/glossary/force-control

# What Is Force Control in Robotics?

**Force control** means a robot joint regulates its **output force/torque** rather than just its position, giving the robot a controllable "softness" when contacting the environment or people. Landing a step, carrying objects with two arms, and human-robot collaboration all depend on it — a position-controlled joint rams through unexpected contact, while a force-controlled joint yields.

## The Three Approaches

- **Torque control**: command joint torque directly — the lowest-level, fastest loop; the `T_ff` feed-forward torque in the [MIT protocol](/glossary/mit-protocol-can) serves exactly this;
- **Impedance control**: command a virtual spring-damper (stiffness Kp, damping Kd) so position error maps to torque — the standard for legged robots; one MIT-protocol frame carries Kp/Kd and torque feed-forward together;
- **Admittance control**: measure external force, output position corrections — common on arms with wrist force sensors, bandwidth-limited by the position loop.

## Do You Need a Torque Sensor?

Friction, efficiency variation, and hysteresis make current-based output-torque estimation less accurate in high-ratio joints, so precision applications often use joint or end-effector force sensors. A low-ratio [quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) joint can estimate output torque from phase current after calibrating motor torque constant, gearbox efficiency, and friction. This supports some proprioceptive and impedance-control tasks, but it does not replace sensing for every high-accuracy or safety-critical application. BXI [85/70/50-series joint actuators](/motors/advanced-motors) support joint control through a 19.5:1 [planetary design](/glossary/planetary-gearbox), MIT-protocol CAN commands, and [dual absolute encoders](/glossary/dual-absolute-encoder).

---

Source: https://www.bxirobotics.com/glossary/frameless-torque-motor

# What Is a Frameless Torque Motor?

A **frameless torque motor** is delivered as **a stator and rotor**, without its own housing, output shaft, or bearings. Both parts integrate directly into the joint structure, reducing redundant packaging and shortening the drivetrain. It is common in [integrated joint actuators](/glossary/joint-motor) and humanoid robot joints.

## Why Joint Modules Use Frameless Motors

- **Structural fusion**: the stator is heat-fitted into the joint housing and the rotor couples straight to the [planetary gearbox](/glossary/planetary-gearbox) sun gear — the shortest possible drivetrain, shrinking joint volume by roughly 30%;
- **Large-diameter pancake form**: torque motors use high pole counts at a large air-gap radius for low-speed, high-torque output — exactly a joint's duty profile, and the source of high [torque density](/glossary/torque-density);
- **Coaxial hollow bore**: the annular stator/rotor geometry naturally supports [hollow-shaft cable routing](/glossary/hollow-shaft-motor) through the joint's center.

## Inrunner vs. Outrunner

Frameless torque motors come as inrunners (rotor inside — better cooling, lower inertia) and outrunners (rotor outside — more torque per volume, flatter). Dynamic legged robots often pick outrunners for torque density. The BXI [85/70/50-series joint actuators](/motors/advanced-motors) build on a frameless torque motor with a 19.5:1 planetary reduction in a [quasi-direct-drive](/glossary/quasi-direct-drive) architecture, completed by [dual absolute encoders](/glossary/dual-absolute-encoder).

---

Source: https://www.bxirobotics.com/glossary/harmonic-drive

# What Is a Harmonic Drive?

A **harmonic drive** (strain wave gear) transmits torque by elastically deforming a thin-walled **flexspline** against a rigid circular spline via an elliptical wave generator, achieving a **50–160:1 reduction in a single stage** with near-zero backlash in a compact, lightweight package. It is one of the three precision robot gearbox families alongside cycloidal (RV) and planetary drives.

## Harmonic Drive vs. Planetary Gearbox

| Dimension | Harmonic drive | [Planetary gearbox](/glossary/planetary-gearbox) |
| --- | --- | --- |
| Single-stage ratio | 50–160:1 | 3–10:1 (stackable) |
| Backlash | Near zero | Small but present |
| Impact tolerance | Flexspline is shock-sensitive | Solid gears, robust |
| Backdrivability | Poor (high friction) | Good at low ratios |
| Typical joints | Cobot wrists/elbows, humanoid arms | Legs, high-dynamic joints |

One-line verdict: **choose harmonic for precision, planetary for impact tolerance and force control**.

## How Robots Choose

Positioning accuracy makes harmonic drives the default in industrial and collaborative robot arm joints. But humanoid and quadruped legs must absorb landing impacts and estimate external forces from motor current, where the high friction and flexspline fatigue of a high-ratio harmonic drive fall short — so legs overwhelmingly use the [quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) planetary route. The BXI [85/70/50-series joint actuators](/motors/advanced-motors) use a 19.5:1 planetary design balancing [torque density](/glossary/torque-density) with backdrivability. Base joints needing extreme torque often use cycloidal (RV) gears — hence the common split: RV at the base, harmonic in the arms, planetary in the legs.

---

Source: https://www.bxirobotics.com/glossary/hollow-shaft-motor

# What Is a Hollow Shaft Motor?

A **hollow shaft motor** is a motor or actuator whose output shaft has a through-bore along its rotation axis. Unlike solid-shaft designs, cables, hydraulic lines, and sensor harnesses can pass straight through the joint's center of rotation instead of looping around the outside.

## Why Robots Need Hollow Shafts

High-DOF robots — such as a 31-DOF humanoid — chain many joints in series. If every joint routes its cables externally:

- harnesses fatigue and fail where joints flex repeatedly;
- exposed cables limit range of motion and snag on the environment;
- wiring complexity and assembly time grow rapidly with joint count.

Routing through the center bore solves all three at once, which is why hollow-shaft designs are central to modern integrated [joint motors](/glossary/joint-motor).

## Engineering Trade-offs

A larger bore passes more wiring but complicates the layout of the rotor, encoders, and gearbox. The BXI [85/70/50-series joint motors](/motors/advanced-motors) provide 6–10 mm hollow bores, using a [planetary gearbox](/glossary/planetary-gearbox) to balance torque density against routing space.

---

Source: https://www.bxirobotics.com/glossary/joint-motor

# What Is a Joint Motor?

A **joint motor** (also called a robot joint module or integrated actuator) packs a frameless torque motor, gearbox, encoders, driver, and bearings into one compact unit. Mounted directly at a robot's joints, it outputs controlled torque and rotation — the "muscle" of humanoid robots, quadrupeds, and robotic arms.

## Core Components

- **Frameless torque motor**: generates raw torque and sets the power ceiling.
- **Gearbox**: trades speed for torque; common choices are the [planetary gearbox](/glossary/planetary-gearbox) and harmonic drive.
- **Encoders**: measure joint angle for closed-loop control; high-end designs use [dual absolute encoders](/glossary/dual-absolute-encoder).
- **Bearing**: carries loads while preserving rotational precision; load-bearing joints typically use a [crossed roller bearing](/glossary/crossed-roller-bearing).
- **Driver**: runs current/velocity/position loops and receives commands over CAN/CANFD buses (see [MIT protocol](/glossary/mit-protocol-can)).

## Key Specifications

The parameters that matter most are **rated torque** (continuous output), **peak torque** (short-term ceiling), weight, envelope dimensions, reduction ratio, and communication interface. A humanoid robot typically needs 20–40 joint motors across several torque tiers — the BXI Elf 3, for example, runs 31 [85/70/50-series joint motors](/motors/advanced-motors) spanning 25–150 N·m peak torque.

For a sizing methodology, see the [joint motor selection guide](/blog/joint-motor-selection-guide).

---

Source: https://www.bxirobotics.com/glossary/mit-protocol-can

# What Is the MIT Protocol?

The **MIT protocol** (MIT Mode) is a CAN-bus motor control protocol originating from the MIT Mini Cheetah open-source legged robot. Its core idea: compress a full control command into one CAN frame carrying five quantities — target position p, target velocity v, position gain Kp, velocity gain Kd, and feed-forward torque τ.

## Hybrid Force-Position Control

The motor computes output torque as `τ_out = Kp·(p − p_actual) + Kd·(v − v_actual) + τ`. By tuning the gains, one protocol covers three control modes:

- **Pure torque control**: Kp = Kd = 0, feed-forward torque only — the basis of high-dynamic force control;
- **Position control**: normal Kp/Kd values track position like a servo;
- **Impedance control**: intermediate gains make the joint behave as a spring-damper for compliant contact.

## Why It Became a De Facto Standard

The protocol is open, compact (one frame per command), and force-control-native. Many [joint motor](/glossary/joint-motor) vendors now support it, creating a cross-brand control interface. All BXI [85/70/50-series joint motors](/motors/advanced-motors) support MIT-protocol CAN/CANFD, and paired with the [PCIE-CAN control module](/motors/control-modules) achieve >1000 Hz whole-robot control loops.

---

Source: https://www.bxirobotics.com/glossary/planetary-gearbox

# What Is a Planetary Gearbox?

A **planetary gearbox** is a gear train built from a central sun gear (input), several planet gears, and an outer ring gear. The planets spin on their own axes while orbiting the sun — like planets around a star, hence the name. It converts a motor's high speed into low-speed, high-torque output.

## Structural Advantages

- **Load sharing**: torque is split across multiple planet gears, giving high torque density and impact tolerance.
- **Coaxial input/output**: the compact cylindrical form factor suits integrated [joint motors](/glossary/joint-motor) naturally.
- **High efficiency**: single-stage efficiency typically exceeds 95%, higher than harmonic drives.
- **Hollow-friendly**: a through-bore at the center enables [hollow-shaft cable routing](/glossary/hollow-shaft-motor).

## Planetary vs. Harmonic

Harmonic drives offer large ratios (50–160) and zero backlash but lower efficiency, limited stiffness, and higher cost. Planetary gearboxes have smaller ratios (roughly 3–10 per stage, ~20 with compounding) and slight backlash, but win on efficiency, shock tolerance, and cost — decisive advantages for high-dynamic humanoid leg joints.

The BXI [85/70/50-series joint motors](/motors/advanced-motors) all use a 19.5-ratio planetary design with a uniform 100 rpm rated output speed, so sizing reduces to picking the right torque tier.

---

Source: https://www.bxirobotics.com/glossary/quasi-direct-drive

# What Is a Quasi-Direct Drive Actuator?

A **quasi-direct drive (QDD) actuator** pairs a large-diameter, high-torque motor with a **low reduction ratio** gearbox (typically 5–20), an approach popularized by legged-robot projects such as MIT Cheetah. "Quasi-direct" means close to direct drive (no reduction) while keeping a small amount of gearing.

## The Key Property: Backdrivability

The higher the reduction ratio, the harder it is to drive the motor backwards from the output (reflected friction and inertia scale with the ratio squared). A low ratio keeps the joint **backdrivable**:

- external impacts are absorbed by the motor yielding rather than the gears fighting them, protecting the drivetrain during landings and collisions;
- output torque can be estimated from motor current after calibrating torque constant, efficiency, and friction, reducing reliance on dedicated joint torque sensors in some applications;
- force-control bandwidth stays high — essential for jumping and running.

## QDD vs. High-Ratio Drives

High-ratio harmonic and cycloidal drives typically provide high output torque and holding capability but have more backdrive resistance. QDD favors dynamic response and compliant control, so it is common in dynamic legged robots. BXI [85/70/50-series joint motors](/motors/advanced-motors) use a 19.5-ratio [planetary design](/glossary/planetary-gearbox) with MIT-protocol control (see [MIT protocol](/glossary/mit-protocol-can)).

---

Source: https://www.bxirobotics.com/glossary/reflected-inertia

# What Is Reflected Inertia?

**Reflected inertia** is the motor rotor's inertia as seen at the gearbox output side: **J_reflected = J_rotor × N²**, where N is the gear ratio. Inertia scales with the **square** of the ratio — a 100:1 joint presents **10,000x** the rotor inertia at its output, while a 10:1 joint presents only 100x.

## Why High-Ratio Joints Feel Stiff and Break Under Impact

Anything pushing on the joint output must also accelerate that squared-up rotor inertia:

- **Poor impact tolerance**: at foot touchdown or collision, the impulse arrives faster than any controller can react, and the energy lands directly on gear teeth and flexsplines — a leading failure mode of high-ratio joints in legs;
- **Hard to backdrive**: reflected inertia plus friction makes the joint nearly immovable by hand, killing passive compliance;
- **Limited force-control bandwidth**: to fake softness, the controller must first fight the huge equivalent inertia, capping active-compliance bandwidth and fidelity.

Quotable one-liner: **double the gear ratio, quadruple the reflected inertia — the physical reason leg joints avoid high ratios.**

## How QDD Keeps Reflected Inertia Low

This is the core trade of the [quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) approach: use a large-diameter, high-[torque-density](/glossary/torque-density) motor so the gear ratio can stay low (typically 6–20:1), cutting reflected inertia by **1–2 orders of magnitude** versus a harmonic solution (50–160:1). The joint becomes naturally backdrivable, impacts are absorbed softly, and external torque can be estimated from motor current alone. BXI's [85/70/50-series joint actuators](/motors/advanced-motors) pick a 19.5:1 planetary ratio as exactly this engineering balance between output torque and reflected inertia.

---

Source: https://www.bxirobotics.com/glossary/robot-data-collection

# What Is Robot Data Collection?

**Robot data collection** (embodied-AI data collection) is the process of synchronously recording paired data — observations (camera images, joint states) plus actions (joint commands) — while a robot performs a task, typically demonstrated by a human via [teleoperation](/glossary/teleoperation). This data is the raw material for training [VLA models](/glossary/vla-model) and imitation-learning policies: what a model can learn is bounded by what was collected.

## The Leader-Follower Teleoperation Pipeline

Leader-follower arms are a common real-robot collection setup:

1. The operator moves a lightweight **leader** arm; the **follower** arm mirrors its joint motion in real time;
2. The system records multi-camera images, follower joint states, and action commands at a fixed rate (typically 30-50 Hz or higher);
3. Each completed task is saved as one trajectory (episode); a task usually needs tens to hundreds of episodes;
4. Data is filtered and aligned, then fed into the training pipeline.

When leader and follower correspond joint-to-joint, demonstration is intuitive and can reduce online kinematic-mapping complexity. The BXI UpperBody 1 [dual-arm teleoperation platform](/robots/robotic-arms) uses a dual-leader, dual-follower architecture and can integrate with ROS2 recording workflows; available interfaces and data formats depend on the project configuration.

## Real-Robot Data vs Simulation Data

- **Real-robot data** captures true friction, deformation, lighting, and sensor noise, so trained policies deploy directly — but collection is labor-intensive and costly;
- **Simulation data** scales cheaply and in parallel, but differs from reality and must be bridged via [Sim-to-Real](/glossary/sim-to-real) techniques.

In practice the two are mixed: simulation provides scale, real data anchors the true distribution.

## How Imitation Learning Consumes the Data

The ALOHA/ACT line of work trains policies directly on leader-follower bimanual trajectories: ACT (Action Chunking with Transformers) predicts a short chunk of future actions as a unit. The required demonstration count depends on task difficulty, data quality, and the desired generalization range. These methods make teleoperated data collection plus imitation learning a practical route for training [embodied-AI](/glossary/embodied-ai) policies.

---

Source: https://www.bxirobotics.com/glossary/rv-reducer

# What Is an RV Reducer?

An **RV reducer** (a member of the **cycloidal** gear family) uses a **two-stage compound structure: a planetary front stage feeding a cycloidal pin-wheel output stage**. Because the cycloidal disc engages many pins simultaneously, it achieves **30–200:1 ratios** with the highest torsional rigidity and shock overload capacity (commonly 2–5x rated) among the three precision reducer families, covering tens to thousands of N·m. The name comes from Nabtesco's RV series and is now generic for the architecture.

## Harmonic vs. RV vs. Planetary: The Three-Way Split

| Dimension | RV reducer | [Harmonic drive](/glossary/harmonic-drive) | [Planetary gearbox](/glossary/planetary-gearbox) |
| --- | --- | --- | --- |
| Single-unit ratio | 30–200:1 | 50–160:1 | 3–10:1 (stackable) |
| Rigidity / shock | High | Flexspline is relatively shock-sensitive | Good |
| Backlash | ~1 arcmin | Near zero | Small but present |
| Weight / size | Large and heavy | Compact | Moderate |
| Backdrivability | Poor | Poor | Good at low ratios |
| Typical joints | Industrial robot base/shoulder | Wrists, humanoid arms | Humanoid/quadruped legs |

A common application pattern is **RV at industrial robot bases, harmonic drives in lightweight arms, and low-ratio planetary gears in backdrivable legs**. The final choice still depends on load, precision, mass, and shock requirements.

## Why RV Reducers Are Less Common in Humanoid Legs

A six-axis industrial robot is bolted to the floor, so its base joint benefits from RV rigidity and load capacity. A humanoid must carry every joint it uses: at comparable torque levels, an RV solution is typically heavier than a low-ratio planetary design, reducing whole-robot [torque density](/glossary/torque-density). Higher friction and lower backdrivability also increase uncertainty in current-based output-torque estimation. Humanoid legs that prioritize low mass and compliant control therefore often use planetary [quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) actuators. BXI's [85/70/50-series joint actuators](/motors/advanced-motors) use a 19.5:1 planetary design with up to 150 N·m peak torque.

---

Source: https://www.bxirobotics.com/glossary/sim-to-real

# What Is Sim-to-Real?

**Sim-to-Real** is the approach of training robot control policies in a physics simulator (such as MuJoCo or Isaac Gym/Lab) and then deploying them on real hardware. Training on real robots is slow, expensive, and breaks hardware; simulation runs thousands of robots in parallel at faster-than-real-time speed, which is what makes reinforcement learning of humanoid walking practical at all.

## The Core Challenge: The Reality Gap

Simulation never matches reality exactly — friction, latency, motor characteristics, and sensor noise all differ. This mismatch is the **sim-to-real gap**. The main mitigations:

- **Domain randomization**: randomly perturb simulation parameters (mass, friction, latency) during training, forcing the policy to become robust;
- **System identification**: measure the real robot's motor response and inertial parameters accurately and write them back into the simulator;
- **Actuator modeling**: model the torque-speed behavior of the [joint motors](/glossary/joint-motor) explicitly — often the deciding factor for legged-robot transfer.

## Hardware Requirements

Deployed policies command joints at hundreds of Hz, demanding high-bandwidth force control (see [quasi-direct drive](/glossary/quasi-direct-drive) and the [MIT protocol](/glossary/mit-protocol-can)) and low-latency buses. The BXI [Elf 3 humanoid](/robots/humanoid-robot) ships with a MuJoCo environment and ROS2 SDK, and its >1000 Hz PCIE-CANFD control architecture supports the full simulation-to-hardware workflow.

---

Source: https://www.bxirobotics.com/glossary/six-axis-force-torque-sensor

# What Is a Six-Axis Force/Torque Sensor?

A **six-axis force/torque sensor** (6-axis F/T sensor) measures **three orthogonal forces (Fx, Fy, Fz) and three torques (Mx, My, Mz)** at one point, typically using strain gauges or capacitive elements. Range, accuracy, and sample rate should be confirmed from the selected sensor's datasheet.

## Where It Goes and What It Does

- **Wrist**: between the arm flange and the tool, for precision assembly, polishing/deburring, and hand-guided teaching — any [force-control](/glossary/force-control) task that needs true end-effector contact forces;
- **Ankle**: on humanoid feet, measuring ground reaction force and center of pressure (ZMP/CoP), a key input for walking balance;
- **Test rigs**: torque calibration of joint actuators, impact testing.

## Do You Still Need One If the Joint Does Current-Based Force Control?

[Quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) joints can estimate output torque from phase current after motor and drivetrain calibration. This can support joint-level compliance, but accuracy depends on friction, temperature, efficiency, and model error. A practical comparison is:

| Scenario | Current-based estimation (QDD) | External 6-axis F/T sensor |
| --- | --- | --- |
| Leg impact absorption, whole-body compliance | Useful control feedback; accuracy depends on calibration | Fit according to control and safety needs |
| Precision assembly | Does not directly measure six-axis tool forces | Usually the better fit |
| Foot ZMP / ground reaction force | Indirect estimate only | Direct and reliable |
| Hardware and integration | No dedicated joint sensor added | Adds sensing, calibration, and cabling |

One-line verdict: **QDD current estimation supports joint-level compliance, while six-axis F/T sensors directly measure tool or foot loads**. They can be combined according to accuracy and safety requirements.

---

Source: https://www.bxirobotics.com/glossary/teleoperation

# What Is Teleoperation?

**Teleoperation** is the technique of a human operator controlling a robot's motion remotely in real time, with the robot executing tasks on site. The operator side (leader) captures human motion intent; the robot side (follower) reproduces it and streams back visual and other feedback, closing a human-in-the-loop control loop.

## Two Kinds of Value

1. **Immediate utility**: for tasks where autonomy isn't ready, teleoperation makes robots useful now — hazardous-environment work, remote assembly, live demonstrations.
2. **Data collection**: the observation–action trajectories produced by teleoperation are **the highest-quality training data for imitation learning**. Today's mainstream [embodied-AI](/glossary/embodied-ai) training pipelines run on large-scale teleoperated demonstrations.

## Common Forms

- **Kinematically-matched leader arms**: a small leader arm mirroring the follower's structure — direct mapping, high precision;
- **VR / motion capture**: headsets and controllers, or full-body mocap, driving whole-body humanoid motion;
- **Shared autonomy**: the human gives high-level commands while the robot handles balance and trajectories.

The BXI [UpperBody 1 dual-arm platform](/robots/robotic-arms) ships with a plug-and-play operator console, and the [Elf 3 humanoid](/robots/humanoid-robot) supports both teleoperation and autonomous modes for manipulation data collection.

---

Source: https://www.bxirobotics.com/glossary/torque-density

# What Is Torque Density?

**Torque density** is the torque an actuator can deliver per kilogram of its own mass, expressed in **N·m/kg**, quoted either at peak or at rated torque. It is an important [robot joint motor](/glossary/joint-motor) metric because actuators mounted in a limb become payload for every joint upstream.

## How to Compute and Compare It

Torque density = output torque ÷ total actuator mass (gearbox and encoders included). Always align the basis before comparing:

- **Peak vs. rated**: marketing figures are usually peak-basis, often 3× or more above the rated-basis number;
- **Driver included or not**: integrated actuator modules include the drive electronics — bare-motor figures are not comparable;
- Current integrated joint actuators typically reach 30–70 N·m/kg peak; top products exceed 100 N·m/kg.

For the BXI [85/70/50 series](/motors/advanced-motors) (peak basis): the [BXI8515-19](/motors/bxi8515-19) delivers 150 N·m ÷ 1.4 kg ≈ **107 N·m/kg**, the [BXI7010-19](/motors/bxi7010-19) 50 N·m ÷ 0.8 kg ≈ 63 N·m/kg, the [BXI5018-19](/motors/bxi5018-19) ≈ 64 N·m/kg, and the [BXI5014-19](/motors/bxi5014-19) 50 N·m/kg.

## Where High Torque Density Comes From

A large-diameter, high-pole-count outrunner torque motor provides the base torque; a [planetary gearbox](/glossary/planetary-gearbox) multiplies it at low added mass. The [quasi-direct-drive (QDD)](/glossary/quasi-direct-drive) approach uses a low-to-medium reduction ratio to balance torque density against backdrivability, while [harmonic drives](/glossary/harmonic-drive) can quote higher density figures at the cost of impact tolerance and backdrivability.

---

Source: https://www.bxirobotics.com/glossary/vla-model

# What Is a VLA Model?

A **VLA model** (Vision-Language-Action model) is an end-to-end robot foundation model: it takes camera images and a natural-language instruction (e.g. "put the cup in the drawer") as input and directly outputs robot actions — joint angles, end-effector poses, or gripper commands. By compressing perception, language understanding, and motion generation into one network, VLA has become one of the most watched policy architectures in [embodied AI](/glossary/embodied-ai).

## Relationship to LLMs

VLA models are typically built on top of vision-language models (VLMs): they inherit the semantic and commonsense knowledge of large language models, then replace or extend the output head with action tokens or continuous actions. A useful mental model: an LLM predicts the next word; a VLA predicts the next action — it is the language-model family extended into the physical world.

## Why Real-Robot Data Is Essential

The internet holds vast text and images, but almost no paired "image + instruction → joint action" data. VLA training therefore depends heavily on real-robot demonstrations, mostly collected via [teleoperation](/glossary/teleoperation). Simulation data can add scale, but must cross the [Sim-to-Real](/glossary/sim-to-real) gap. Data diversity — across tasks, scenes, and embodiments — often matters more for generalization than parameter count.

## Representative Work

- **RT-2** (Google DeepMind): co-trains a VLM with robot action tokens, demonstrating that web-scale knowledge transfers to manipulation;
- **OpenVLA**: an open-source VLA trained on large open robot datasets, designed for community fine-tuning;
- **π0** (Physical Intelligence): uses flow matching to generate continuous actions, targeting a general cross-embodiment manipulation policy.

Deploying and fine-tuning VLA on real hardware requires a platform that executes high-rate action commands reliably — BXI's [humanoid robot](/robots/humanoid-robot) and [dual-arm platform](/robots/robotic-arms) both expose ROS2 interfaces suited to VLA research.

---

Source: https://www.bxirobotics.com/glossary/world-model

# What Is a World Model?

In robotics, a **world model** is an internal predictive model of environment dynamics learned by an agent: given the current observation (images, state) and an intended action, it predicts what the environment will look like next. With a world model, a robot can roll out the consequences of actions "in its head" — imagine first, act second — instead of trial-and-erroring every step in the real world.

## Relationship to VLA and Reinforcement Learning

- **Versus [VLA models](/glossary/vla-model)**: a VLA maps observations and instructions directly to actions — a reactive policy. A world model explicitly captures "action → consequence" and supports planning and rollouts. The two are complementary: a world model gives a VLA the ability to imagine and verify.
- **Versus reinforcement learning**: in model-based RL, the policy trains largely by "dreaming" inside the learned world model, requiring far fewer real interactions and cutting the cost of real-robot sampling. This parallels training in a simulator followed by [Sim-to-Real](/glossary/sim-to-real) transfer — a world model is essentially a differentiable "neural simulator" learned from data.

## Why It Is a 2026 Embodied-AI Focus

First, progress in video generation shows large models can absorb substantial physical regularities from massive video corpora — a foundation for general world models. Second, real-robot data is scarce and expensive, so the field wants world models to amplify limited [teleoperation](/glossary/teleoperation) data by generating and evaluating additional virtual experience. Third, [embodied AI](/glossary/embodied-ai) has high safety stakes: rehearsing inside an internal model before executing on hardware is a natural way to reduce risk.

Whatever the technical route, world models must ultimately be validated on real hardware. BXI's [humanoid robot](/robots/humanoid-robot) ships with MuJoCo simulation and a ROS2 SDK, supporting the full loop of in-model training followed by real-robot validation.