Materials and Sensors

1. Introduction

1.1. Background

The MIIT Guiding Opinions on the Innovation and Development of Humanoid Robots (2023-11) lists the "machine body" (lightweight skeleton, high-strength body, high-precision sensing) as a key technology cluster, and in Column 2 gives the research directions for the sensor family: high-precision bionic eyes, wide-bandwidth bionic hearing, high-resolution multi-point-contact human-like e-skin, and bionic smell — with e-skin being listed as a key national product research direction. In the Harness context, this family defines the physical ceiling of what an agent can "see, touch, and know about its own posture."

On the market side, the six-axis force sensor is the segment with the richest data in this direction: in 2025, domestic brands held a 39.1% overall share of the Chinese six-axis force sensor market (GGII basis), while the MIR Rui Industry basis gives 58.8%, overtaking foreign brands for the first time — the two bases are listed side by side because their statistical scopes differ; the domestic substitution rate in the humanoid-robot segment exceeds 90%. Ramp-up of complete machines drives upstream demand: the Agibot Ling G2 achieves sub-millimeter force-controlled assembly with 100% automotive-grade components and joint torque sensors, and the Yuanzheng A2-W is equipped with a 360° LiDAR and 6 depth cameras — sensing and materials have "gone on board" at scale.

1.2. Definition and Scope

The materials-and-sensors track covers robots' perceptual input and physical body materials. In the AI Harness context it refers specifically to the following stages:

Scope notes: This track does not cover algorithmic processing of sensing data (belongs to Brain (VLA), Cerebellum (Motion Control)); the mechanical part of actuator-integrated torque sensors is discussed in Actuators & Servo, while this track covers their calibration and detection.

1.3. Position in the AI Harness System

The mapping of materials and sensors onto the six-layer model is as follows:

Core judgment: The bottleneck of the materials-and-sensors track is at L1. Sensing data quality determines the upper bound of what the model "sees": leading six-axis force accuracy of 0.1% FS, 10 kHz sampling, and crosstalk no greater than 0.3% are the hard constraints on current perceptual input. The industry expects leading accuracy of 0.05% FS in 2026 and 0.01% FS in 2028, with unit price dropping from USD 800–1500 to below USD 300 (industry association basis) — each leap in accuracy and cost simultaneously raises L1 context quality and L5 observation credibility. This is also the fundamental reason "force sensing lands before vision": force data can directly enter the control loop (L5 observation → L2 execution), whereas visual semantics still require the brain to digest.

1.4. Current State of Development

Six-axis force sensors. Three technology routes: strain-gauge (mainstream, 76.95% share, decoupling accuracy 0.5% FS, temperature drift ±0.02% FS/℃, 10 kHz sampling), piezoelectric (0.1 ms-level response, unit price above USD 5000), and capacitive. 2025 humanoid-robot segment shares (GGII basis): Landian 72.6%, Kunwei 25.7%, Yuli Instrument 6.9%; smart-robot overall shipment basis (MIR): Kunwei at 53%, industry No. 1 — the two bases have different statistical scopes and must be listed side by side. Miniaturization benchmarks: Yuli Instrument M3701F1 with a 6 mm diameter, 1 g weight, and nonlinearity no greater than 0.5% F.S; Xinjingcheng with the world's smallest diameter of 9.5 mm.

Six-axis force sensor selection criteria table. The lateral comparison table below uses publicly verifiable parameters only (fields not disclosed by manufacturers are marked [To be filled]; it is forbidden to extrapolate from same-category products; share and shipment bases are given above and in Figure 4-1):

Most manufacturers' public materials generally do not give the range field, marked [To be filled]; on the price side, the industry trend basis is: unit price is expected to drop from USD 800–1500 to below USD 300 (industry association basis). When selecting, use this table's fields as the comparison framework; the accuracy, crosstalk, and response fields directly determine the L1 context quality ceiling (see the core judgment in Section 1.3).

Figure 4-1 | Dual-basis comparison of six-axis force sensor shares

Tactile sensing and e-skin. Arrayed tactile sensing skin (E-Skin) has left the laboratory: it covers fingertips and palms and can detect contact forces at the micro-newton level, enabling dexterous grasping of fragile objects such as eggs and glass cups (industry review basis); Tanyuan Sensing's flexible tactile sensor offers a 0–100 N range with accuracy up to 0.01 N and a million-cycle lifetime (exhibitor materials basis). On the complete-machine side, Tesla Optimus uses a multi-layer tactile sensing system covering most of the hand surface (perceiving texture, temperature, and pressure, with a silicone protective layer balancing sensitivity and durability); Fourier GR-3 arranges 31 touch sensors across the head and torso.

Comparison of key tactile and e-skin solution points. The deployment positions and key parameters of four representative solutions are compared laterally below:

The four solutions fall into two distinct commitment paths — "device-level accuracy" (Tanyuan Sensing) and "complete-machine-level coverage" (Optimus, GR-3). When selecting, first clarify the deployment position and perception goal, then compare the accuracy and cycle-life fields.

IMU and inertial sensing. An IMU measures acceleration and angular velocity to reconstruct motion trajectories; it is easy to connect and not constrained by space, but has low accuracy and no absolute position accuracy. Optical motion capture is highly accurate (it can output independent information for each bone segment) but requires large venues and has poor occlusion resistance. Tesla Optimus uses Xsens full-body inertial motion capture to collect data. Public accuracy specs for the IMUs built into humanoid robots (bias stability, range) are generally not disclosed by manufacturers, marked [To be filled].

Lightweight materials. Material comparison: magnesium alloy 1.7–1.8 g/cm³ (joint housings, shells, skeletons; magnesium-to-aluminum price ratio 0.87); PEEK 1.3 g/cm³ (specific strength 8× that of aluminum alloy, continuous service temperature 260℃, about RMB 300,000/ton); nylon 1.15–1.2 g/cm³; carbon fiber 1.5–1.6 g/cm³ (tensile strength above 3500 MPa, specific strength 16× that of steel); aluminum alloy about 2.7 g/cm³ (cost-effectiveness baseline); titanium alloy 4.5 g/cm³.

2. Glossary of Terms

3. Case Studies

3.1. Landian: Leader in Both Share and Parameters for Humanoid-Robot Six-Axis Force Sensors

Background. Landian was founded in 2019 with ties to the Aerospace Science and Technology Group background. In the humanoid-robot six-axis force sensor segment, its 2025 share is 72.6% (GGII basis, aggregated through the Sensitive Components and Sensors Branch of the China Electronic Components Industry Association), ranking first; its domestic shipments of joint torque sensors account for over 95%, with 70,000+ units shipped in H1 2025 (company/association basis).

Approach. Product parameters: accuracy 0.1% FS, response frequency 10 kHz, overload capacity 300% (another source gives 500%; the two bases are listed side by side), crosstalk no greater than 0.3%, volume reduced 90%, weight reduced 80%. The product line covers six-axis force sensors and joint torque sensors, of which the joint torque sensor is the typical form of "actuator-integrated sensing" — the Ling G2's sub-millimeter force-controlled assembly relies on joint-level torque feedback.

Outcome. A position that is close to "first in both share and parameters" in the humanoid segment shows that force sensing is now a confirmed requirement for humanoid-robot mass production: multi-joint per unit plus distributed wrist placement make six-axis force/torque sensing the fastest-ramping incremental component. Note: the share data comes from the GGII basis aggregated by the association (medium-high credibility), and the overload capacity parameter has two bases — 300% and 500% — which must be listed side by side when cited. Credibility: medium-high.

3.2. Kunwei Technology: Lead Drafter of the National Standard and Calibration Methodology

Background. Kunwei Technology was founded in 2018 with a background in aerospace research institutions, and is the core lead drafter of GB/T 43199-2023 Specification for Testing of Multi-dimensional Force/Torque Sensors for Robots — the first domestic testing standard for robot force sensors. Under the MIR Rui Industry basis, Kunwei's 2025 shipment share in China's smart-robot field (humanoid + collaborative) is 53%, ranking first in the industry; its market share in the collaborative-robot segment exceeds 70%.

Approach. Product parameters: accuracy better than 0.1% FS, accuracy better than 0.3% FS (World Robot Conference exhibitor materials basis); six-axis joint loading calibration is the core national-standard method; annual capacity reached 60,000 units by end of June 2026.

Outcome. The value of the Kunwei case lies not in a single parameter but in the combined position of "standard + capacity": leading the drafting of the national standard means its calibration method becomes the industry-wide arbiter (equivalent to writing the L5 evaluation rules into the national standard), while the 60,000-unit annual capacity corresponds to the sensor supporting capability for humanoid robots landing at the ten-thousand-unit scale. The basis difference must be emphasized: MIR's 53% (humanoid + collaborative overall) and GGII's 25.7% (humanoid segment) have different statistical scopes — the two bases are listed side by side and do not constitute a contradiction. Credibility: medium-high.

3.3. Lightweight Materials: From Weight-Reduction Numbers to Battery-Life Gains

Background. The humanoid robot is a battery-powered whole-body motion system, so weight reduction translates directly into battery life and joint longevity. Industry compilation basis: a 40% weight reduction can raise battery life from 2 hours to 6 hours (another basis: "a 10 kg weight reduction raises battery life by 20%–30% and reduces joint wear by 15%;" the two bases are listed side by side); PEEK usage per humanoid robot is about 6.5–8 kg (another basis: "about 15% of total weight"), and carbon fiber usage is 8–10 kg.

Approach and outcome. Complete-machine evidence chain: Unitree G1 uses titanium-alloy–carbon-fiber composite joints, with a whole-machine weight of 35 kg (official basis); UBTech Walker X's magnesium-alloy gearbox reduces weight by 55% and noise by 12 dB; Tesla Optimus Gen2 reduces whole-machine weight by 10 kg (magnesium alloy + PEEK/CF, media basis), and the Gen3 knee support structure uses magnesium alloy to reduce weight by 42% (TechkTimes basis). Component evidence chain: Boston Dynamics Atlas's CFRP joint bracket is 45% lighter than aluminum alloy, with a flexural modulus of 230 GPa; Kemeng Innovation's PEEK-composite harmonic reducer cuts overall weight by 61% and raises the torque-to-weight ratio by 74%; the PEEK reverse planetary roller screw reduces material waste by 90% compared with metal CNC. Supply-chain side: Zhongyan (the domestic PEEK resin leader) is the PEEK resin supplier for Tesla Optimus joint bearings, with a capacity of 1000 tons expanding to 5000 tons in phase two (commissioned 2026-09); Zhongfu Shenying's T1200-grade carbon fiber engineering-grade sample has a strength of 7566 MPa; the 1X Neo Gamma uses braided nylon for its shell.

Process-consistency risk and testing basis. All of the weight-reduction figures above are the joint output of "material + process": replacing the process route for the same material can significantly shrink the delivered performance, and this is the risk most easily underestimated in lightweight material selection. The three component-level evidence chains in the current search material are all joint outputs of material and process:

• PEEK route: Kemeng Innovation's PEEK-composite harmonic reducer cuts overall weight by 61% and raises the torque-to-weight ratio by 74%; the PEEK reverse planetary roller screw reduces material waste by 90% compared with metal CNC — the delivery of both metrics depends on batch consistency in the molding process, not on PEEK's intrinsic material properties by themselves.
• Carbon fiber route: Boston Dynamics Atlas's CFRP joint bracket is 45% lighter than aluminum alloy, with a flexural modulus of 230 GPa — the layup scheme and curing process directly determine how much of the modulus is delivered.
• Magnesium alloy route: UBTech Walker X's magnesium-alloy gearbox reduces weight by 55% and noise by 12 dB — porosity control in the die-casting process is a prerequisite for fatigue life.

Testing basis: material-level indicators (density, specific strength, temperature resistance) only prove that the material is qualified; component-level delivery must be verified back through joint- or whole-machine-level tests — joint-module performance can be compared against GB/T 43200-2023 Performance and Test Methods for Robotic Integrated Joints; on the supply side, batch stability during the capacity ramp-up must also be watched (e.g., batch management after Zhongyan's phase-two 5000-ton PEEK capacity comes online).

Conclusion. Lightweighting is not "swapping in a lighter material" but "selecting materials by location": titanium-alloy–carbon-fiber for load-bearing joints, magnesium alloy for housings and gearboxes, PEEK for transmission parts, and nylon for shells — every choice is a joint optimization of density, specific strength, temperature resistance, and cost. Credibility: whole-machine figures are from official or media bases (medium-high); the weight-reduction–battery-life conversion is an industry compilation basis (medium, two bases listed side by side).

4. Practice Standards

4.1. AGENTS.md Norms (Materials and Sensors Track)

Below is the complete, copy-ready AGENTS.md for the materials-and-sensors track — a trimmed and strengthened version of the group-level AGENTS.md:

# AGENTS.md —— 具身智能组 · 材料与传感器方向

## 角色与边界
- **角色**：材料与传感器方向工程智能体，负责六维力与触觉传感器选型、标定方案编制、IMU 与动捕方案对比、轻量化材料选型与传感数据分析。
- **边界**：不直接操作六维力标定台与材料试验设备；不出具标定证书（证书由计量资质人员签发）。
- **第一原则**：精度声明必须有标定依据。未标定数据不得用于任何安全相关结论。

## 环境假设
- 声明传感链路：六维力传感器型号、标定状态与有效期（是否按 GB/T 43199-2023 方法标校）、采样率、安装方式。
- 声明材料工况：载荷谱、温度范围、耐化学与磨损要求、目标寿命。
- 声明数据合规状态：动捕与触觉数据涉及人体的须有知情同意记录。

## 上下文加载顺序（Context Budget）
- 必载：任务判据、传感标定状态表、材料工况需求。
- 次载：候选件规格摘要、份额与价格数据（含口径标注）。
- 禁止：原始高速传感数据流（10 kHz 级）进入上下文，以统计摘要代替。

## 工具契约
- 传感对比表字段统一：量程、精度（% FS）、准度、串扰、响应频率、抗过载、重量、尺寸、价格。
- 材料对比表字段统一：密度、比强度、耐温、成本、典型部位。
- 份额类数据必须双口径并列（GGII 人形细分 vs MIR 智能机器人整体），禁止择一。
- 参数两口径冲突（如蓝点触控抗过载 300% 与 500%）并列标注。

## 任务执行流程（SOP）
- S1 解析感知或减重需求为可判定指标；S2 规格对表与口径核对；S3 编制标定或试验方案（引用 GB/T 43199-2023 等标准）；S4 数据回收与漂移分析（设备由人工操作）；S5 产出选型建议与安全影响评估。

## 验证与证据要求
- 传感数据用于安全结论（PFL、失稳检测）前必须确认标定有效期内。
- 减重-续航换算标注口径与来源性质；两口径并存时不合并。
- 材料性能引用密度、比强度等基础物理量与官方数据，工艺性声明标注厂商口径。

## 失败与升级策略
- 传感数据漂移或标定过期：冻结相关任务，安排重新标定。
- 材料试验与规格不符：升级供应商澄清，禁止单方下调裕度。

## 安全与合规红线
- 安全回路信号源（力觉、触觉）的标定与有效性检查不可跳过。
- 涉人体数据（动捕、触觉）脱敏与知情同意是入库前置条件。

## 禁止事项
- 禁止用未标定数据支撑安全结论；禁止编造传感精度与材料参数；禁止口径选择性引用。

## 输出格式
- 报告结构：需求 → 对比表（统一字段、口径标注）→ 方案 → 数据分析 → 建议与安全影响。

## 评估与自检
- 自检项：标定状态确认、口径并列、单位规范、安全影响评估、[待填写] 规范。

4.2. SKILL.md Norms (Materials and Sensors Track)

---
name: embodied-material-sensor
description: 材料与传感器方向技能。当需要选型六维力/触觉/惯性传感器、编制标定方案、分析传感数据质量，或做轻量化材料选型对比时使用。
version: 1.0
created: 2026-09-12
---

# 材料与传感器方向技能

## 适用场景
- 六维力传感器选型与标定方案编制（GB/T 43199-2023）。
- 触觉皮肤与电子皮肤方案对比；IMU 与动捕方案对比。
- 轻量化材料选型（镁合金 / PEEK / 碳纤维 / 钛合金）与减重收益分析。

## 前置条件
- 感知或减重需求可量化；标定设备与有效期记录可查。
- 涉人体数据有知情同意记录。

## 输入
- 需求指标表；候选件规格与来源；标定状态表；工况载荷谱。

## 输出
- 统一字段对比表、标定/试验方案、数据分析报告、选型建议与安全影响评估。

## 执行步骤
1. 需求解析：转为量程、精度、响应、寿命等可判定指标。
2. 对比表：统一字段；份额与参数多口径并列。
3. 方案编制：引用国标方法；设备操作归人工。
4. 数据分析：漂移、温漂、串扰核查；安全相关数据核对标定有效期。
5. 结论：建议 + 口径标注 + 安全影响。

## 质量标准（DoD）
- 精度声明有标定依据；口径并列；单位规范；涉人体数据合规；无占位符。

## 常见失败与处理
- 数据漂移：先查标定有效期再查温漂与安装应力。
- 材料试验不符：升级供应商澄清，不单方降裕度。

## 示例
- 为腕部选六维力传感器：以 0.1% FS 精度、10 kHz 响应、串扰不超过 0.3% 为判据对表，按 GB/T 43199-2023 拟定六维联合加载标校方案。

4.3. Implementation Checklist

5. Summary

The Harness logic of the materials-and-sensors track is the "most physical" of the four tracks in this group: L1 context engineering in the software domain discusses "which tokens to show the model," while this track discusses "what the sensors can physically let the model see" — 0.1% FS force-sensing accuracy, a 10 kHz sampling rate, and micro-newton-level tactile thresholds are the hard ceiling of this link.

Three engineering conclusions. First, force sensing is the category in the sensor family to first reach scale: the domestic substitution rate exceeding 90% in the humanoid segment, the dual-No. 1 pattern of Landian at 72.6% and Kunwei at 53% on the MIR basis, and the standardization of calibration methods under GB/T 43199-2023 show that force sensing has completed the full path of "laboratory → standard → capacity"; whereas built-in IMU accuracy remains a blind spot in public information, and e-skin is still at the "leaving the laboratory" stage — the maturity gradient within the family is itself a priority basis for selection. Second, basis discipline for share data is the most critical in this track: GGII (humanoid segment) and MIR (smart-robot overall) can give opposite No. 1 results, so any statement of "who is No. 1 in China" without stating a basis is an invalid claim. Third, weight reduction is a cross-layer benefit: it simultaneously improves battery life (the energy budget of L2 execution), joint longevity (L4 ledger), and dynamic performance (L3 control margin), which explains why location-based selection of magnesium alloy, PEEK, and carbon fiber has become the consistent practice among leading complete machines.

Information Gap Statement

The following items have not been confirmed by primary sources and have been flagged in the body text:

1. Public accuracy specs for the IMUs built into humanoid robots (bias stability, range, model): generally not disclosed on manufacturers' complete-machine spec pages, marked [To be filled].
2. Landian overload capacity parameter: two bases coexist at 300% and 500% (inconsistency within industry-association aggregated materials), already marked side by side.
3. Overall domestic share of six-axis force sensing: 39.1% (GGII basis) and 58.8% (MIR basis) have different statistical scopes; listed side by side without merging.
4. Weight-reduction–battery-life conversion: a 40% weight reduction raises battery life from 2 h to 6 h, and "a 10 kg weight reduction raises battery life by 20%–30%" — two bases listed side by side; per-unit PEEK usage of 6.5–8 kg and "about 15% of total weight" are two bases listed side by side.
5. Tesla Optimus tactile system details (silicone protective layer thickness, tendon-cable force-feedback implementation): media basis; the official spec has not been fully released.
6. Planetary roller screw life improvement: 3–5× and 15× are two bases listed side by side, not choosing either.
7. Tanyuan Sensing flexible tactile 0.01 N accuracy and million-cycle lifetime: exhibitor materials basis; not reproduced by a third party.

6. References

1. Domestic well-known manufacturers of six-axis 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
2. Domestic substitution inside a single sensor — Tencent News, 2026-08-10. https://new.qq.com/rain/a/20260810A0BX2G00
3. Official exhibitor profiles from the World Robot Conference (Kunwei Technology, Landian, Tanyuan Sensing, etc.) — World Robot Conference, 2025. https://www.worldrobotconference.com/news/3235.html
4. Guiding Opinions on the Innovation and Development of Humanoid Robots — Ministry of Industry and Information Technology, 2023-11. https://www.ncsti.gov.cn/zcfg/zcwj/202311/P020231103482413965397.pdf
5. Report on China issuing 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
6. Robot data collection methods: teleoperation and motion capture, driven by both real and synthetic data — Shenwan Hongyuan (republished by San Ge Pijiang Report), 2025-05. https://www.sgpjbg.com/labels/jiqirenshujucaijifangshi/1/6757805.html
7. Unitree official support page (G1 complete-machine specifications) — Unitree. https://support.unitree.com/home
8. Report on the launch of Fourier GR-3 — Securities Times, 2025-08-06. https://www.stcn.com/article/detail/2975614.html
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 robotics industry — Zhonghong Net, 2026-06. https://www.zhonghongwang.com/show-278-464160-1.html
11. Specification for Testing of Multi-dimensional Force/Torque Sensors for Robots (GB/T 43199-2023) — national standard, National Technical Committee for Robot Standardization (SAC/TC 591); URL not confirmed, cited by title + standard number.
12. Performance and Test Methods for Robotic Integrated Joints (GB/T 43200-2023) — national standard, National Technical Committee for Robot Standardization (SAC/TC 591); URL not confirmed, cited by title + standard number.
13. Panoramic analysis of robot motor technology — Future Think Tank, 2025-12. https://www.industrysourcing.cn/article/471909