Cerebellum

1. Introduction

1.1. Background

The MIIT's Guiding Opinions on the Innovative Development of Humanoid Robots (2023-11), Column 1, defines the “key technology cluster of the robot cerebellum” as: high-fidelity system modeling and simulation, multi-body dynamics modeling with online behavior control, typical biomimetic locomotion behavior representation, and autonomous learning of whole-body coordinated movement; the target capabilities are robust whole-body coordinated locomotion in unstructured environments, dexterous manipulation, and human-robot interaction. This policy statement delineates the four research objects of this direction: modeling and simulation (the training ground), dynamics control (WBC), behavior representation (gait and skill library), and autonomous learning (reinforcement learning).

On the industry side, the consensus given by the 2025 industry review is: the brain is responsible for decomposing natural-language instructions into logical steps, while the cerebellum runs on a local computing platform and, at millisecond-level frequency, combines IMU, encoder, and force-sensor feedback to output motor control signals (in real time and disturbance-resistant). At the prospectus level, Unitree has explicitly built a full-stack “body + cerebellum + brain” technical system (as reported by China Economic Net); its terminology is consistent with the policy language — “cerebellum” has already moved from an academic concept to the business-segment description of a listed company.

1.2. Definition and Scope

Cerebellum direction covers the real-time motion control layer of robots; in the AI Harness context, it specifically refers to the following links:

Boundary note: this direction only covers the real-time control and action generation layer; semantic planning belongs to Brain (VLA), and actuator physical parameters belong to Actuators & Servo. The System 1 portion of the dual-system VLA (real-time action generation) belongs to this direction.

1.3. Positioning in the AI Harness System

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

Core judgment: the bottleneck of the cerebellum direction lies in L5 (the sim2real domain gap). Both sides of the quantitative evidence are equally clear: on the positive side, synthetic data brings a +40% performance improvement (per NVIDIA's official claim) and Galaxy General achieves a 95% grasping success rate with pure simulation training (per company claim); on the negative side, a policy that converges in simulation may still become unstable on the real robot due to dynamics differences, latency, and sensor noise. All 27 HumanoidBench tasks are on the simulation side; there is still no unified comparable benchmark for real-robot motion performance — this is exactly the biggest gap in the L5 layer of this direction.

1.4. Current Status

Frequency anchors. The publicly disclosed frequency parameters of the large-brain/small-brain separation architecture are as follows. It must be emphasized: these frequencies come from different subsystems of different models and cannot be compared directly; they serve only as reference anchors for each layer's latency budget.

Figure 2-1 | Frequency comparison of the dual large-brain/small-brain systems (schematic scale)

Schematic: drawn based on the terminology of Section 1.4 of this document.

Simulation toolchain. The Newton open-source physics engine (co-built by NVIDIA + Google DeepMind + Disney Research, based on Warp, compatible with MuJoCo and Isaac Lab); MuJoCo-Warp is expected to accelerate by more than 70×; the mainstream simulation and synthesis tools are Isaac, MuJoCo, and Genesis. HumanoidBench (UC Berkeley, 2024-03) is the first whole-body humanoid simulation evaluation benchmark: 27 tasks (15 manipulation + 12 locomotion), covering whole-body coordinated control and dexterous-hand fine manipulation, making it the benchmark of choice for L5 evaluation on the cerebellum side.

Engineering selection discipline for the simulation toolchain. The four mainstream toolchains have different applicability boundaries; mixing them and directly comparing across engines is a common source of unreproducible sim2real results:

Selection discipline: the engine version and downgrade path must be written into the environment declaration (corresponding to the “tool contract” in the AGENTS.md of Section 4.1 and item 8 “engine declaration” of the Section 4.3 checklist); evaluation results across engines must not be compared directly — the difference in scores of the same policy under different engines is itself part of the domain gap, and must be recorded as an independent variable rather than cherry-picked for citation.

Data side. ACT-class approaches, where 10 minutes of real-robot teleoperation data can reach an 80% success rate (per Shenwan research report terminology), coexist with a trend toward scaling: NVIDIA EgoScale achieves One-shot generalization using more than 20,000 hours of first-person data. The data pyramid (web video → synthetic → real-robot teleoperation) is the general framework for sim2real data governance.

2. Glossary

3. Case Studies

3.1. Unitree H1 / G1: Dual Anchors of Motion Control Parameters and Mass Production Scale

Background. Unitree Technology is the representative of the “hardware-first body school” and the domestic manufacturer that discloses the fullest motion control parameters for the cerebellum. Based on the prospectus reporting (China Economic Net, 2026-06): 2023–2025 revenue of 1.59 / 3.93 / 16.99 亿元 (hundred million yuan); over 5,500 humanoid robots shipped in 2025, No. 1 globally, with a production-sales ratio of 91.24%; listed on the STAR Market on 2026-08-19, becoming the “first humanoid-robot stock on the A-share market”. Note: the 2025 net profit excluding non-recurring items of 5.91 亿元 (hundred million yuan) and the net profit attributable to shareholders of 2.78 亿元 (hundred million yuan) are listed side by side in two terminologies and must be distinguished when citing.

Approach. The H1 has 19 degrees of freedom in total (5 per leg, 4 per arm, 1 at the waist); the in-house M107 joint motor has a maximum torque of 360 N·m; the joint torque distribution is ~360 N·m at the knee, ~220 N·m at the hip, ~59 N·m at the ankle, and ~75 N·m at the arm (official terminology); a locomotion speed of 3.3 m/s (a world record), with potential motion performance above 5 m/s; the motion-control compute unit is an i5-1235U (dedicated) + i7 (development). The G1 is 132 cm tall and weighs 35 kg (base version); the base version has 23 degrees of freedom, and the EDU version allows 23–43 degrees of freedom to be configured (the JD.com retail page lists a separate 29-DOF terminology; all three terminologies coexist, with the official one authoritative); a maximum joint torque of 120 N·m, a walking speed of 3 km/h, and a price starting at 9.9 万元 (ten thousand yuan).

Results. With a 35 kg whole-robot weight and a price in the ten-thousand-yuan range, the G1 lowers the verification bar for whole-body motion control to a level affordable by university laboratories, becoming one of the most commonly used domestic platforms in RL sim2real research; the H1's 3.3 m/s locomotion speed demonstrates the performance ceiling of the model-predictive + whole-body control route. At the 2025-08 World Humanoid Robot Games, Team Tiangong won the 4×100 m relay final and Star Dynamics won the standing high jump and set a world record (official self-reported terminology); motion-performance competition has become a public showcase for cerebellum capabilities. Credibility: the whole-robot parameters are official terminology (high); the prospectus figures are media-reported (medium-high).

G1 EDU version tiers and research ecosystem fit. The G1's version tiers are a layered design aimed at the research ecosystem: the base version has 23 DOFs, weighs 35 kg, and is priced from 9.9 万元 (ten thousand yuan); the EDU version's DOFs can be configured between 23–43, the EDU U2 advanced version weighs 37 kg, and the JD.com retail price range is ¥84,500–¥209,000.

The low-tier version covers gait and balance verification, while the high-end EDU version supports whole-body coordinated control (WBC) and dexterous-hand collaboration research with more degrees of freedom; three DOF terminologies coexist (official 23 base / 23–43 configurable / e-commerce 29), and citations follow the official one with the parallel terminologies noted. The “affordable platform + configurable DOF tiers” and the “affordable for university laboratories” judgment above are mutually causal, jointly explaining the G1's platform-side ecological niche in the RL sim2real research chain.

3.2. NVIDIA GR00T N1 System 1: Real-Time Action Generation Amplified by Synthetic Data

Background. NVIDIA released Isaac GR00T N1 at GTC 2025 (2025-03-18), officially calling it the “world's first open, fully customizable humanoid robot foundation model”. NVIDIA does not build robots; instead it outputs “simulation training ground + foundation model + synthetic data pipeline” as public infrastructure for both the cerebellum and the brain — this itself is a Harness-form productization.

Approach. GR00T N1 adopts a dual-system architecture: System 2 is a VLM (NVIDIA-Eagle + SmolLM-1.7B) for reasoning and planning; System 1 is a Diffusion Transformer that translates System 2's plans into precise continuous actions, is trained on human demonstrations and Omniverse synthetic data, and is co-fine-tuned with System 2. On the data side, a three-tier pyramid is used (web video → Omniverse synthetic data → real-robot teleoperation data); the Isaac GR00T Blueprint uses Cosmos Transfer to generate 780,000 synthetic trajectories in 11 hours (equivalent to roughly 6,500 hours, i.e., 9 consecutive months, of human demonstrations).

Results. Synthetic + real data improves performance by 40% over purely real data (official); with the full dataset, the average real-robot task success rate is 76.8%. Adapted platforms include Fourier GR-1 and 1X NEO Gamma; ecosystem partners include Agility, Boston Dynamics, Mentee, and NEURA. This case also provides a quantitative methodology for sim2real: the gain from synthetic data is not an adjective but two reconcilable numbers, 40% and 76.8%. Credibility: official terminology (high).

3.3. AgiBot: Expedition A2's Sim-to-Real Live Demo and Lingjing G2's Automotive-Grade Force Control

Background. AgiBot Robotics is the domestic representative of the “full-stack hardware + software school”: the 1,000th general-purpose embodied robot rolled off the line on 2025-01-06, and the 10,000th rolled off on 2026-03-30 (Expedition A3); on 2025-11-20, Expedition A2 set the Guinness World Record for the “longest distance walked by a humanoid robot”. Its value on the cerebellum side lies in moving sim2real from paper metrics into public live venues and automotive-grade mass-production scenarios.

Approach. The IROS 2025 live demo “from sim to real, from RL to IL”: relying on simulation training data, Expedition A2 completes the full process of desktop-item recognition and cleaning without any advance tuning in the actual scene (per Zhidx reporting). The Lingjing G2 (released 2025-10-16) has 26 DOFs throughout the body, 100% automotive-grade components, IP42 protection, sub-millimeter force-control assembly precision, two 7-DOF industrial arms, and a 5 kg rated payload, equipped with joint torque sensors and a 3-DOF waist — building force-control feedback into mass-produced hardware.

Results. The “no advance tuning” live demo is landmark evidence of sim2real engineering: simulation is the cerebellum's regression suite and the real robot is the delivery site, making the transfer success rate between them a field-verifiable metric. The Lingjing G2's automotive-grade components and sub-millimeter force control show that the ceiling of the cerebellum's control precision is being set by hardware (joint torque sensors) rather than by algorithms alone. On the open-source side, the AgiBot World million-trajectory real-robot dataset (over 1 million trajectories, 217 tasks, 100 homogeneous robots) and the Lingqu OS together provide L4 data assets to the industry. Credibility: medium-high (official releases and media reports).

Lingqu OS and AgiBot World's L4 data-asset operations. The other half of AgiBot's cerebellum capability sits at L4 — data and runtime assets:

• AgiBot World (open-sourced 2024-12): over 1 million real-robot trajectories covering 217 tasks across five scenarios, collected from 100 homogeneous robots; in 2025-02 it was accompanied by the open-source simulation framework AgiBot Digital World.
• Lingqu OS and AimRT: wrap the above data assets into a schedulable robot operating system and middleware runtime.
• Genie Studio: an agent-based deployment platform that serves as the entry point for policy iteration and on-robot deployment.

The value of collecting from a homogeneous robot fleet lies in data consistency: the million trajectories produced on the same platform with the same sensor layout can serve directly as the “regression dataset” for simulation training and sim2real evaluation, corresponding to the Harness L4 mapping of “real-robot trajectory dataset ≈ long-term memory / artifact library”. The “platform shipment → data feedback → policy iteration” data flywheel is thus closed — this is the structural advantage that distinguishes full-stack vendors from purely algorithm-based routes, and it is also consistent with the “teleoperation and real-robot data mapping to L4” terminology in this group's 01-brain.md.

4. Practice Standards

4.1. AGENTS.md Specification (Cerebellum Direction)

The following is the complete, copy-ready AGENTS.md for the cerebellum direction, a trimmed and reinforced version of the group-level AGENTS.md:

# AGENTS.md —— 具身智能组 · 小脑方向

## 角色与边界
- **角色**：小脑方向工程智能体，负责运动控制与 WBC 算法开发、强化学习训练与 sim2real 迁移、轨迹生成与仿真评测。
- **边界**：不修改关节限位表与功率力限制（PFL）参数；真机下发只产出待人工确认的指令包。
- **第一原则**：仿真先行。任何策略未经仿真回归通过，不得进入真机流程。

## 环境假设
- 声明物理引擎与版本（Isaac Lab / MuJoCo / Genesis / Newton）、域随机化参数表（质量、摩擦、延迟、传感噪声）。
- 声明本体构型与自由度（如 H1 19 自由度 / G1 基础版 23 自由度）、关节限位表、控制频率预算（50–200 Hz）。
- 声明评测基准版本（HumanoidBench 27 任务 / 项目内回归集）与力矩、速度上限。

## 上下文加载顺序（Context Budget）
- 必载：任务判据、限位表、当前策略版本与回归集最好成绩。
- 次载：域随机化配置、失败案例（跌倒、超限）清单。
- 禁止：原始 1 kHz 关节反馈流、完整 URDF 进入上下文。

## 工具契约
- 训练与仿真脚本版本锁定、随机种子固定、同配置可复现。
- 仿真结果必须含退出码、日志路径、成功率与力矩峰值、引擎版本。
- 引擎降级（如 Isaac Lab 不可用改用 MuJoCo）须在结论中标注"开源引擎结果，不得作为放行依据"。

## 任务执行流程（SOP）
- S1 解析任务；S2 固定环境复现基线；S3 仿真训练（多种子，检查力矩与速度不超限位）；S4 HumanoidBench 与项目内回归集双评测；S5 影子验证；S6 真机分级试跑（低限幅起步、逐级放开、每级人工放行）；S7 失败案例入回归集。

## 验证与证据要求
- 仿真成绩须报多种子统计（均值与最差种子），禁止单种子最优值。
- sim2real 声明必须给出真机验收环境与成功率；缺真机数据须声明"仅仿真"。
- 引用第三方整机参数标注官方 / 第三方口径（如 G1 自由度三口径并列）。

## 失败与升级策略
- 仿真连续两次调参不收敛：升级人工分析。
- 真机失稳、超限：立即急停，冻结任务，只做数据分析。

## 安全与合规红线
- 力矩与速度限幅不可修改；任务需求冲突走人工评审。
- 真机操作人在回路；运动性能展示（速度、跳跃）须在隔离场地进行。

## 禁止事项
- 禁止用仿真成绩冒充真机成绩；禁止修改限幅"硬闯"目标工况。

## 输出格式
- 报告结构：形态声明 → 仿真结果（种子与回归集版本）→ 真机结果（干预记录）→ 结论与遗留风险。

## 评估与自检
- 自检项：多种子统计、限幅合规、双评测、影子验证记录、失败案例入回归集。

4.2. SKILL.md Specification (Cerebellum Direction)

---
name: embodied-cerebellum-sim2real
description: 小脑方向技能。当需要开发运动控制或 WBC 算法、训练强化学习策略、执行 sim2real 迁移与轨迹生成评测，并要求多种子可复现与真机分级验证时使用。
version: 1.0
created: 2026-09-12
---

# 小脑方向技能

## 适用场景
- 步态与全身控制（WBC）算法开发与调优。
- 强化学习训练（Isaac Lab / MuJoCo / Genesis）与 sim2real 迁移。
- 轨迹生成评测（VLA 动作头输出质量、HumanoidBench 27 任务）。

## 前置条件
- 物理引擎版本与域随机化参数表已锁定；限位表与 PFL 参数已确认。
- 评测基准版本固定；真机任务有监督人员与急停。

## 输入
- 任务目标与成功判据；本体 MJCF/URDF 与限位表；奖励函数或参考轨迹。

## 输出
- 策略权重与检查点、多种子评测报告、真机试验安全记录。

## 执行步骤
1. 复现基线策略在回归集上的成绩。
2. 训练：多种子并行，记录训练曲线与力矩峰值。
3. 评测：HumanoidBench 与项目内回归集双报告。
4. sim2real：影子验证 → 低限幅试跑 → 逐级放开（每级人工放行）。
5. 失败案例全部入回归集，产出遗留风险清单。

## 质量标准（DoD）
- 多种子统计（均值 + 最差种子）；限幅未越界；双评测齐备；真机记录完整；口径标注。

## 常见失败与处理
- sim2real 失败：扩域随机化覆盖，补真机失败场景数据，禁改限幅。
- 训练发散：固定种子最小复现，分离奖励设计与数据问题。

## 示例
- 在 Isaac Lab 训练双足平地行走策略，HumanoidBench 行走任务组评测后于 G1 上低限幅试跑。

4.3. Implementation Checklist

5. Summary

The cerebellum direction is the “hard real-time-ification” of the Harness Loop in the physical world: the tool-execution loop of the software domain responds to text and code on the order of seconds, while the cerebellum responds to IMU, encoder, and force-sensor feedback at a frequency of 50–200 Hz — the layer positions (L2 execution + L3 real-time loop) have not changed, but the latency budget has been compressed by four orders of magnitude, and the cost of an error has shifted from “running it once more” to “dropping a machine”.

Three engineering conclusions. First, the dual-system frequency layering is an actionable design guideline: the public anchors of 7–9 Hz for the semantic layer, ~50 Hz for the action head, and 200 Hz for joint control show that “slow thinking, fast execution” is not a metaphor but a frequency budget that can be written into interface contracts. Second, simulation is already the primary production environment of the cerebellum: GR00T N1's +40% synthetic-data gain, AgiBot Expedition A2's “no advance tuning” live demo, and Galaxy General's 95% grasping with pure simulation (company claim) together push the sim2real transfer success rate into the position of an acceptance metric; but the Galaxy General and Games results indicate by their terminology that the real-robot side still lacks a unified benchmark. Third, hardware is redrawing the algorithm ceiling: Unitree's M107 at 360 N·m and the Lingjing G2's automotive-grade joint torque sensors show that the L2 boundary of the cerebellum is determined by actuators and sensing — this is exactly why this group's four directions must be read together.

Information Gap Statement

The following items have not been confirmed by first-hand sources and have been annotated in the text:

1. GR00T N1 System 1's output frequency: no official figure has been published, not listed in Figure 2-1, [To be filled].
2. Unified benchmark for real-robot motion performance: HumanoidBench covers the simulation side, and there is no cross-vendor comparable benchmark for real-robot gait and whole-body control, [To be filled].
3. The original URL of Unitree's prospectus was not obtained: figures such as revenue, shipments, and production-sales ratio come from media reports by China Economic Net / China Youth Daily Net, etc.; the net profit attributable to shareholders (2.78 亿元/hundred million yuan) and the net profit excluding non-recurring items (5.91 亿元/hundred million yuan) are listed side by side without being merged.
4. The G1's three DOF terminologies (official 23 base / 23–43 configurable / e-commerce 29) coexist without choosing one.
5. Star Dynamics STAR 1's 55 DOFs: a re-quoted figure compiled by Huachuang, not officially confirmed, and not cited in this document.
6. Some of AgiBot's Lingjing G2 release times and parameters differ across materials from different points in time; the official release note prevails.

6. References

1. Unitree official support page (H1 / G1 / M107 motor specifications) — Unitree Technology. https://support.unitree.com/home
2. NVIDIA Releases Isaac GR00T N1 — NVIDIA Newsroom, 2025-03-18. https://nvidianews.nvidia.com/news/nvidia-releases-isaac-gr00t-n1-worlds-first-open-humanoid-robot-foundation-model-and-simulation-libraries
3. Accelerate Generalist Humanoid Robot Development with Isaac GR00T N1 — NVIDIA Developer Blog, 2025. https://developer.nvidia.com/blog/accelerate-generalist-humanoid-robot-development-with-nvidia-isaac-gr00t-n1/
4. Guiding Opinions on the Innovative Development of Humanoid Robots — Ministry of Industry and Information Technology, 2023-11. https://www.ncsti.gov.cn/zcfg/zcwj/202311/P020231103482413965397.pdf
5. Report: China releases a humanoid robot and embodied intelligence standards system (2026 edition) — People's Daily Overseas Edition, 2026-02-28. https://peoplesdaily.pdnews.cn/china/er/30051524844
6. LIBERO benchmark horizontal comparison table (arXiv 2602.18224) — arXiv, 2026-02. https://arxiv.org/pdf/2602.18224
7. ALOHA 2 / Mobile ALOHA entry — aiwiki, 2024–2025. https://aiwiki.ai/wiki/aloha_2
8. Robot data collection methods: teleoperation and motion capture, real and synthetic data dual-driven — Shenwan Hongyuan (re-quoted by Sange Pijiang Report), 2025-05. https://www.sgpjbg.com/labels/jiqirenshujucaijifangshi/1/6757805.html
9. Fourier GR-3 release report — Securities Times, 2025-08-06. https://www.stcn.com/article/detail/2975614.html
10. World Robot Conference official exhibitor introduction — World Robot Conference, 2025. https://www.worldrobotconference.com/news/3235.html
11. Embodied intelligence accelerates in depth — Digital China Net, 2025-12. https://www.digitalchina.gov.cn/2025/xwzx/szkx/202512/t20251230_5263761.htm
12. Comprehensive analysis of robot motor technology — Future Think Tank, 2025-12. https://www.industrysourcing.cn/article/471909
13. 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/
14. Embodied intelligence: decoding the breakthrough path of China's robot industry — Zhonghong Net, 2026-06. https://www.zhonghongwang.com/show-278-464160-1.html