Servo Motors and Actuators

1. Introduction

1.1. Background

Actuators are the physical carrier of embodied intelligence's "tool execution" capability. In terms of policy, the Ministry of Industry and Information Technology's Guidelines on Innovation and Development of Humanoid Robots (2023-11) lists "machine limbs" (humanoid robotic arms, dexterous hands, legs and feet) and the "machine body" (lightweight skeleton, high-strength chassis) as key technology groups, and explicitly calls for the deployment of high-power-density actuators. In market terms, actuators are the segment with the most certain volume growth: a single humanoid robot needs 14–17 harmonic reducers (industrial robots need only 3–5, per Xueqiu supply-chain analysis); a dexterous hand accounts for roughly 20%–30% of a humanoid robot's total cost, and Dongwu Securities forecasts the dexterous-hand market to reach 552 billion yuan by 2030 (forecast figures). CIC (Zhishu Consulting) forecasts China's robotic harmonic-reducer shipments to grow from about 1.4 million units in 2025 to 20.7 million units by 2030 (CAGR 72.4%, with the humanoid share rising from 11% to 82%).

The 2025 mass-production figures validate the scale of this segment: Unitree shipped over 5,500 humanoid robots in 2025 (prospectus-reported figures) with a self-developed/self-manufactured rate for core components exceeding 90%; AgileX's (Zhiyuan) 10,000th general-purpose embodied robot rolled off the line on 2026-03-30. The volume growth of the chassis directly drives upstream components such as joint motors, reducers, ball screws, and coreless motors.

1.2. Definition and Scope

Servo Motors and Actuators covers the drive, transmission, and grasping of robot joints and end effectors. In the AI Harness context, it specifically refers to the following segments:

Boundary note: This direction does not cover control algorithms (attributed to Cerebellum (Motion Control)) or sensing-element principles (attributed to Materials & Sensors); torque sensors built into joints are discussed at the interface between the two directions.

1.3. Position in the AI Harness System

The mapping of actuators in the six-layer model is as follows:

Core judgment: The bottleneck of the actuator direction lies in L2. Once an actuator is finalized, torque density, backlash, and response bandwidth become the ceiling for upper-layer algorithms — no matter how high the cerebellum's control bandwidth, it cannot break through the mechanical resonance of a harmonic reducer. Every improvement at the hardware layer is equivalent to expanding the capability boundary of the L2 tool contract: Leadrive's third harmonics compress transmission precision from the 30–60 arcseconds of traditional second harmonics to within 10 arcseconds BW and improve torsional stiffness by 2–3× (company annual-report/research-report figures). Such an improvement is no less valuable to upper-layer control than an algorithmic iteration.

1.4. Current State of Development

Three-way classification of actuators (Future Think Tank Panoramic Analysis of Robot Motor Technology, 2025-12):

• Flexible actuators: motor + high-reduction-ratio gearbox, high torque density, but with mechanical resonance and low force transparency.
• Elastic actuators: flexible actuator + elastic element, strong impact resistance, but low force-control bandwidth.
• Quasi-direct-drive actuators (QDD): high-torque-density motor + low reduction ratio, high force transparency and strong back-drivability, but low torque density.

Motor topologies: inner-rotor radial flux is mainstream for 30–80 mm joints; the outer rotor is used in 100–180 mm leg-and-foot joints (Unitree, BEMO, and Lingzu Times adopt the "outer rotor + planetary gearbox" quasi-direct-drive mode); axial flux offers high torque density but has technical bottlenecks ("Qinglong" and GAC GoMate have attempted it); coreless motors (8–20 mm) are dedicated to micro joints and dexterous hands. Four core performance requirements: high torque density, low torque ripple, strong overload capability, and heat dissipation.

Figure 3-1｜Actuator scheme comparison matrix (qualitative)

Illustration: drawn based on the industry-research figures in section 1.4 of this document; the red box marks the mainstream leg-and-foot joint route noted by this group.

2. Glossary

3. Case Studies

3.1. Tesla Optimus Gen-3 Dexterous Hand: a composite scheme of tendons + linkages + micro ball screws

Background. Optimus is the whole-machine project receiving the most attention in the actuator direction. Musk has said the dexterous hand "accounts for about half of Optimus's engineering effort." Note that Tesla has never officially released the machine's full specifications; the parameters in this case all come from broker research reports and media (CITIC Securities, RobotToday, WebProNews, etc.) with medium-to-high reliability, and citations must be marked as "third-party figures."

Approach. The Gen-3 dexterous hand has 22 degrees of freedom: 4 each for the index, middle, and ring fingers and 5 each for the thumb and pinky (1 each for DIP/PIP + 2 for MCP, with the pinky including two opposition joints). The drive scheme is a composite of "tendon (flexion/extension) + linkage (abduction/adduction) + micro ball screw": the drivers are placed in the forearm with tendon transmission, and the micro ball screws increase from 13 to 17 per hand; each tendon is paired with a force-feedback sensor. The tendon scheme's freedom advantage (20+) and its lifetime shortfall (5k–10k cycles) both manifest in this case; another supply-chain-rumor figure gives the Gen-3's fault-free operating time at about 500 hours, compared with the industrial-grade standard of 2,000 hours (medium-to-low reliability, for reference only). Value-added figures: the Optimus Gen1 uses 12 coreless motors per hand and 24 across both hands, with a per-unit value of about ¥7200; coreless motors account for about 38% of the dexterous-hand BOM; the dexterous-hand BOM target is to fall from $12,000 to below $3,000.

Impact. The Gen-3 prototype was finalized in October–November 2025 (supply-chain figures); a third-party compilation (humanoid.press, medium-to-low reliability) states the Optimus 3 whole machine is 173 cm, 57 kg, with about 40+ degrees of freedom, a 2.3 kWh battery, 8–10 hours of endurance, and "1,000+ units already internally deployed" (not officially confirmed by Tesla). The confirmable engineering conclusion is that the composite tendon + screw transmission route trades some lifetime for 22 degrees of hand dexterity, with force feedback pushed down to each tendon — which is equivalent to sinking L6's force-boundary sensing into the transmission chain. Reliability: the structural scheme is cross-verified across multiple sources (medium-high); the whole-machine figures are medium-to-low (third party).

3.2. Leadrive Harmonic: third-harmonic reducers and domestic substitution

Background. Harmonic reducers have long been dominated by Harmonic Drive, with three widely divergent global-share figures: about 40% (narrow/robot figures), 58% (2022 figures), and over 80% (distributor-promotional figures); a separate figure of "Chinese vendors' 2025 global share of about 75%" is suspected to mean "Chinese vendors' share of global capacity" — all four figures are presented side by side, and this document does not single out one as the sole truth. Leadrive is the domestic leader, with a domestic market share ranging from 26%–60% under different figures and roughly 12–15% globally.

Approach. Leadrive's self-developed P-type tooth profile breaks Harmonic Drive's IH-tooth-profile patent; its third-harmonic reducer achieves transmission precision within 10 arcseconds (traditional second harmonic: 30–60 arcseconds) and improves torsional stiffness by 2–3×; the Y series offers 1-arcmin precision and a 10,000–12,000-hour lifetime. Commercially: pricing is about 30%–70% of Harmonic Drive's comparable models (about ¥1320–1400 vs. ¥3000–5000); delivery lead time is 1–2 months vs. 3–6 months for imports; capacity is about 600,000 units/year and expanding to 1.4 million. In miniaturization, Laifu Harmonic's 03-series micro reducer weighs only 13 g, targeting dexterous hands and micro joints.

Impact. The third harmonic's 10-arcsecond precision and 2–3× torsional stiffness directly raise the hardware ceiling of cerebellum control — the improvement in precision and stiffness is equivalent to reducing the compensation burden of the control loop. The price and lead-time advantages (30%–70% pricing, 1–2 months lead time) simultaneously compress the humanoid's whole-machine BOM and its production-ramp cycle. Harmonic Drive's moat remains: transmission error below 1 arcmin and the V-Ti-Nb micro-alloyed flexspline material process (industry-material figures). This case is the most direct evidence of "actuator improvement = L2 tool-contract upgrade." Reliability: technical parameters are company annual-report/research-report figures (medium-high); market-share figures conflict across sources (medium, already marked side by side).

3.3. Fourier FSA 2.0: integrated actuators and delivery to ten thousand units

Background. Founded in 2015 and starting with rehabilitation robots, Fourier had cumulatively delivered over 10,000 units by 2025Q1 (Securities Times figures), with a 2025 delivery target of 300 bipedal humanoids — one of the few domestic vendors delivering in the unit of "ten thousand," whose self-developed actuator route has been validated over a long cycle in rehabilitation scenarios.

Approach. The GR-2 carries the self-developed FSA 2.0 actuator (Force Sharing Actuator, an integrated-joint actuator) and a 12-degree-of-freedom dexterous hand; the GR-3 (released 2025-08) is 1.65 m tall, weighs 71 kg, has 55 degrees of freedom across the body, and a 3 kg single-hand load, with dual hot-swappable batteries (3-hour endurance) and 31 touch sensors on the head and torso, continuing to use the self-developed integrated actuator + 12-degree-of-freedom dexterous hand. The FSA route packs the motor, reducer, encoder, and drive-and-control into a single standardized module — that is, the "standard-part-ization" of the L2 tool contract.

Impact. The value of the integrated actuator lies in turning the actuator from a "custom part" into a "standard part with a specification contract": whole-machine integrators select by torque, precision, and response rather than jointly tuning the transmission chain. This is structurally analogous to the software domain's "tool registration + interface contract" Harness approach. The supporting GB/T 43200-2023 Robotic Integrated Joints — Performance and Test Methods provides a national-level arbiter for such modules. Reliability: official and Securities Times figures (medium-high).

4. Practice Standards

4.1. AGENTS.md Specification (Actuator Direction)

The following is the complete, copyable content of the AGENTS.md for the servo motor and actuator direction, an adapted and reinforced version of the group-level AGENTS.md:

# AGENTS.md —— 具身智能组 · 伺服电机与执行器方向

## 角色与边界
- **角色**：执行器方向工程智能体，负责关节电机与减速器选型对比、传动方案分析、台架试验方案编制与试验数据分析。
- **边界**：不直接驱动台架与试验设备（六维力标定台、扭矩台、耐久台为人工操作设备）；不带电带载启动试验。
- **第一原则**：参数必须溯源。选型结论只基于官方规格页或标准试验方法数据。

## 环境假设
- 声明目标关节与工况（峰值扭矩、转速、负载谱、空间包络、散热条件）。
- 声明候选执行器清单与官方规格来源；第三方汇编参数标注"第三方口径，未经官方确认"。
- 声明试验依据标准：GB/T 43200-2023（一体化关节）、GB/T 44251-2024（腿式机器人）按需引用。

## 上下文加载顺序（Context Budget）
- 必载：工况需求表、候选件规格书摘要、在用执行器的衰减台账。
- 次载：供应链口径数据（价格、交期、产能）与来源性质。
- 禁止：完整 CAD 装配体与长扫描报告进入上下文。

## 工具契约
- 对比表字段统一：额定/峰值扭矩、精度（弧秒或 arcmin）、背隙、效率、寿命、重量、价格、交期。
- 数值缺失一律写 [待填写]；禁止用同类产品数据填充。
- 市场份额类数据必须列全冲突口径（如哈默纳科 40% / 58% / 80% 与中国厂商 75% 四口径并列）。

## 任务执行流程（SOP）
- S1 解析工况为可判定指标；S2 规格对表初筛；S3 拟定台架试验方案（依据 GB/T 43200-2023）；S4 试验数据回收与分析（人工执行试验）；S5 产出选型建议与风险清单。

## 验证与证据要求
- 每项关键参数标注来源：官方规格页 / 标准试验 / 研报 / 媒体传闻（传闻不得进入选型结论）。
- 寿命与精度结论必须来自台架实测或标准方法，禁止引用宣传口径。

## 失败与升级策略
- 规格数据冲突：并列呈现并请求人工裁决；禁止静默择一。
- 台架数据与规格书不符：冻结选型，升级供应商澄清流程。

## 安全与合规红线
- 过载保护与机械限位设计不做省略；试验方案须含防护与急停要求（参照 ISO 10218-1/2:2025 安全要求框架）。
- 关键份额与独家供应风险须在结论中显式提示。

## 禁止事项
- 禁止编造规格参数；禁止把供应链传闻写成确认订单；禁止选择性引用有利口径。

## 输出格式
- 选型报告结构：工况 → 候选对比表（字段统一）→ 试验方案 → 数据分析 → 建议与风险（含供应链口径提示）。

## 评估与自检
- 自检项：参数溯源、口径并列、[待填写] 规范、试验依据标准、风险清单齐备。

4.2. SKILL.md Specification (Actuator Direction)

---
name: embodied-actuator-selection
description: 执行器方向技能。当需要对比关节电机、减速器、丝杠、灵巧手等执行器方案，编制台架试验方案，或分析传动精度与寿命数据时使用。
version: 1.0
created: 2026-09-12
---

# 执行器方向技能

## 适用场景
- 关节电机与减速器选型对比（谐波 / RV / 行星 / 丝杠 / 腱绳）。
- 灵巧手驱动方案分析（空心杯 / 无刷有齿槽 / 腱绳 + 连杆 + 微型丝杠复合）。
- 一体化关节台架试验方案编制（GB/T 43200-2023）与试验数据分析。

## 前置条件
- 目标关节工况（扭矩、转速、负载谱、包络）明确。
- 候选件官方规格来源可获取；份额类数据备好多口径。

## 输入
- 工况需求表；候选件规格书；衰减台账；试验设备能力说明。

## 输出
- 统一字段对比表、台架试验方案、数据分析报告、选型建议与风险清单。

## 执行步骤
1. 工况解析：把需求转为可判定指标（含安全裕度）。
2. 规格对表：统一字段初筛；缺失项写 [待填写]。
3. 试验方案：引用 GB/T 43200-2023 拟定项目与判据。
4. 数据分析：实测与规格对照，异常升级澄清。
5. 结论：建议 + 供应链风险 + 口径标注。

## 质量标准（DoD）
- 全部参数带单位与来源；冲突口径并列；传闻不进结论；无占位符。

## 常见失败与处理
- 规格冲突：并列呈现请求人工裁决。
- 试验与规格不符：冻结选型走供应商澄清。

## 示例
- 为髋关节对比两款谐波减速器：以峰值扭矩、10 弧秒级精度、背隙、寿命字段对表，按 GB/T 43200-2023 拟定台架方案。

4.3. Landing Checklist

5. Summary

The Harness logic of the actuator direction can be summarized in one sentence: the actuator is the physical form of the L2 tool contract. In the software domain, the tool contract specifies "what input this tool accepts, what output it promises, and what happens beyond the boundary"; in the actuator domain, this set of promises is translated into rated torque, transmission precision, backlash, efficiency, and lifetime — and once installed on a machine, this can no longer be modified through version updates.

Three engineering conclusions. First, the route debate has no endgame; the working condition decides the selection: the high torque density of flexible actuators, the force transparency of QDD, and the degrees-of-freedom advantage of tendons each carry a cost. Tesla trades a tendon + screw composite for 22 degrees of freedom, and Unitree trades an outer rotor + planetaries for leg-and-foot quasi-direct drive, showing that the right question is not "which route wins" but "what is this joint's load spectrum and control demand." Second, domestic substitution is rewriting the price curve and the precision curve at the same time: Leadrive's third-harmonic 10-arcsecond precision, 30%–70% pricing, and 1–2 month lead time turn the actuator from an "imported custom part" into a "domestic standard part" — the supply-chain precondition for whole machines to move from prototypes to ten-thousand-unit mass production (Unitree 5,500 units, AgileX 10,000 units). Third, standards are the arbiter of the actuator: GB/T 43200-2023 gives integrated-joint performance a national-level test method, turning "actuator specifications ≈ tool contract" from an analogy into an executable acceptance process.

Information-Gap Statement

The following items have not been confirmed by a first-hand source and have been marked in the body of this document:

1. The specific usage and precision of RV reducers (Nabtesco) in humanoid robots: no reliable data was found in this round, [To be filled].
2. Tesla Optimus's official full specifications (degrees of freedom, torque, camera configuration): Tesla has never officially released them; third-party compilations such as humanoid.press are of medium-to-low reliability, for reference only and marked as such.
3. The Gen-3's fault-free operating time of about 500 hours and "1,000+ units internally deployed": supply-chain rumors and third-party claims with medium-to-low reliability.
4. The four harmonic-reducer global-share figures (Harmonic Drive about 40% / 58% / over 80%, Chinese vendors 75%) diverge sharply; they are only presented side by side, with the statistical-figure differences noted.
5. Tuopu Group's roughly 3 billion yuan rotating-joint order and its planned annual capacity of 1 million units: rumor-level figures, not cited by this document as fact.
6. Galbot's six-degree-of-freedom manipulation accuracy error of under 0.5 mm: company/encyclopedia figures, lacking third-party reproduction.
7. The original URL of Unitree's prospectus was not obtained; figures such as the over-90% self-developed/self-manufactured rate come via media retelling.

6. References

1. Panoramic Analysis of Robot Motor Technology — Future Think Tank, 2025-12. https://www.industrysourcing.cn/article/471909
2. Unitree's official support page (H1 / G1 / M107 motor specifications) — Unitree. https://support.unitree.com/home
3. Fourier GR-3 launch report — Securities Times, 2025-08-06. https://www.stcn.com/article/detail/2975614.html
4. Official exhibitor introduction at the World Robot Conference (Inspire Robots RH5DG2, etc.) — World Robot Conference, 2025. https://www.worldrobotconference.com/news/3235.html
5. Guidelines on Innovation and Development of Humanoid Robots — Ministry of Industry and Information Technology, 2023-11. https://www.ncsti.gov.cn/zcfg/zcwj/202311/P020231103482413965397.pdf
6. Report on China's release of the humanoid-robot and embodied-intelligence standard system (2026 edition) — People's Daily Overseas Edition, 2026-02-28. https://peoplesdaily.pdnews.cn/china/er/30051524844
7. Collaborative Robot Safety Standards 2026: ISO 10218-2025 — EVS Int, 2025. https://www.evsint.com/zh-CN/collaborative-robot-safety-standards-2026-iso-10218-2025-ts-15066/
8. Collaborative Robot Safety Standards — InMotion Global. https://www.inmotion.global/resources/cobot-safety/collaborative-robot-safety-standards/
9. Embodied Intelligence Accelerates in Depth — Digital China, 2025-12. https://www.digitalchina.gov.cn/2025/xwzx/szkx/202512/t20251230_5263761.htm
10. Embodied Intelligence: Decoding the Breakthrough Path of China's Robot Industry — Zhonghong Net, 2026-06. https://www.zhonghongwang.com/show-278-464160-1.html
11. Robotic Integrated Joints — Performance and Test Methods (GB/T 43200-2023) — national standard, National Technical Committee on Robotics Standardization (SAC/TC 591), URL unconfirmed; cited by title + standard number.
12. Specification for Inspection of Robot Multi-Dimensional Force/Torque Sensors (GB/T 43199-2023) — national standard, National Technical Committee on Robotics Standardization (SAC/TC 591), URL unconfirmed; cited by title + standard number.
13. Well-Known Domestic Manufacturers of Six-Dimensional Force Sensors for Industrial Robots — Sensitive Components and Sensors Branch, China Electronic Components Industry Association, 2026-03-10. http://sensor.ic-ceca.org.cn/hangyezixun/508.html