Embodied Intelligence Group

1. Introduction

1.1. Positioning of the Group

The Embodied Intelligence Group covers the complete technology stack for agents entering the physical world, divided into four directions: Brain (01), Cerebellum (02), Servo Motors & Actuators (03), Materials & Sensors (04).

The most authoritative policy basis for this division is the Ministry of Industry and Information Technology's Guiding Opinions on the Innovative Development of Humanoid Robots (2023-11): the official framework divides the key technologies of humanoid robots into three technology clusters — "Brain, Cerebellum, and Body" — where the Brain corresponds to LLM-based environment perception, behavior control, and human-robot interaction; the Cerebellum corresponds to the motion-control algorithm library and simulation training environment; the machine limb corresponds to humanoid robotic arms, dexterous hands, and legs/feet; and the machine body corresponds to lightweight skeletons, high-strength bodies, and high-precision sensing. The Humanoid Robot and Embodied Intelligence Standards System (2026 Edition), released 2026-02-28, further covers the entire industrial chain through six major sectors: "Foundational Commonalities, Brain-like & Intelligent Computing, Limbs & Components, Complete Machines & Systems, Applications, and Safety & Ethics." The three-way classification of Turing Award laureate Andrew Yao (body, cerebellum, brain) provides a more accessible framework that is isomorphic to this partition.

Within the AI Harness system, this group offers the strongest explanatory fit for the six-layer model while also incurring the highest constraint cost. The agent runtime in the physical world shares the same structure as the agent runtime in the software domain — goal assembly, planning, tool execution, observation feedback, state update — but the cost of an error in each loop cycle shifts from "rolling back one deployment" to "damaging a device or injuring a person." This group's Harness must therefore upgrade the L6 governance & safety layer from a "recommendation" to a "hard constraint": simulation first, real-machine human-in-the-loop, and a prohibition on autonomously executing irreversible physical actions.

1.2. Core Assertions of the Group

Assertion 1: The "brain/cerebellum separation architecture" is the mainstream consensus of the industry today and the direct mapping of the Harness six-layer model onto the physical world.

• Figure Helix (2025-02) is the first publicly described "dual-system" VLA: the slow system (System 2) runs an onboard VLM at 7–9 Hz, handling scene and language understanding; the fast system (System 1) outputs full upper-body control at 200 Hz. These are the most frequently cited public parameters (official figures) for the "brain/cerebellum separation, frequency-tiered" architecture.
• NVIDIA Isaac GR00T N1 (2025-03-18) adopts an isomorphic dual system: System 2 performs VLM reasoning and planning, and System 1 is a Diffusion Transformer generating continuous actions (official figures).
• The reverse route also exists: Galaxy General (银河通用) has developed its own "Galaxy Astral Brain" (AstraBrain), claiming to be the "world's first whole-body, whole-hand end-to-end embodied foundation model that integrates brain-cerebellum-neural control in a single model," representing the "brain/cerebellum unification" route (company figures). Both routes are presented side by side; this document does not presuppose a single conclusion.
• The 2025 industry review summarizes: the brain is responsible for decomposing natural-language instructions into logical steps (generalization), while the cerebellum runs on a local computing platform, combining IMU/encoder/force-sensor feedback at millisecond-level frequency to output motor control signals (real-time, disturbance-resistant).

Assertion 2: The four directions each anchor to different layers of the Harness six layers, composing a complete "physical agent runtime."

Assertion 3: This group's L6 red line is fundamentally different from that of the software engineering group. In the software domain, the cost of errors is availability and data; in the physical domain, the cost of errors is device damage and personal safety. ISO 10218:2025 (released 2025-02, effective 2025-04-01) has incorporated the collaborative-application safety requirements of ISO/TS 15066 (including Power and Force Limiting, PFL) into its body, becoming the anchor standard for this group's L6 layer. Hence this group's AGENTS.md hard constraints are "simulation first + real-machine human-in-the-loop," rather than the software domain's "sandbox + approval."

Assertion 4: The coincidental overlap of grading numbers must be explicitly distinguished. The group standard T/CIE 298-2025 Humanoid Robot Intelligence Grading adopts a "four dimensions, five levels" scheme (perception & cognition, decision & learning, collaboration & interaction, execution & performance as four dimensions; L1 basic execution level through L5 general intelligence level). This coincidentally shares the L1–L5/L1–L6 numbering with the Harness six-layer model but has completely different meanings, so the group's writing must not conflate the two.

Figure 1-1 | Embodied intelligence full-series stack and Harness six-layer mapping

Data sources and mapping basis: see the in-figure annotations and Section 1.2 of this document; the frequency anchors are from the official statements of Figure / Physical Intelligence / NVIDIA, information as of 2026-09-12.

1.3. Industry Landscape Data

Note: The three shipment perspectives come from different sources and statistical methods; consistent with convention, this document presents them in parallel rather than merging them. On player structure, industry reports classify players into three types: full-stack software-and-hardware players (Figure, AgiBot), hardware-heavy "body school" (Unitree, Agility, Apptronik), and software-heavy "brain school" (Physical Intelligence, Field AI, Galaxy General) — these three happen to echo the two ends of this group's four directions (brain vs. actuators).

2. Glossary

3. Positioning Comparison of the Four Directions

3.1. Comparison Table of Direction Overviews

3.2. Mapping Table of the Four Directions to the Harness Six Layers

3.3. Core Bottleneck Layer of Each Direction

• Brain: the bottleneck is at L5. There is a structural gap between 94%–97% on LIBERO and true open-world success rates — π0.5's core selling point is precisely "completing cleaning tasks in real homes never seen during training," showing that the evaluation loop has not yet caught up with the generalization goal. In addition, official vendor figures for VLA on-device inference latency are largely missing (only the Helix 200 Hz and π0 50 Hz figures can be cited), leaving L2 real-time validation without a unified yardstick.
• Cerebellum: the bottleneck is at L5 (sim2real domain gap). While synthetic data brings a +40% performance gain, the domain gap remains the primary cause of transfer failure; 27-task benchmarks such as HumanoidBench cover only the simulation side, and real-machine motion metrics still lack a unified benchmark.
• Actuators: the bottleneck is at L2. Actuator performance is the physical upper bound of the "tool contract" — once torque density, backlash, and response bandwidth are fixed, no upper-layer algorithm can surpass them. Hardware improvements such as a 2–3× increase in torsional stiffness (Green's third-harmonic) effectively expand the capability boundary of L2 tools.
• Materials & Sensors: the bottleneck is at L1. Sensor data quality determines the upper bound of what the model "sees": leading six-axis force accuracy of 0.1% FS and 10 kHz sampling are the hard constraints on current perceptual input; the industry expects 0.05% FS by 2026 and 0.01% FS by 2028, and every accuracy leap directly raises the ceiling of both L1 and L5.

4. Common Challenges and Countermeasures

5. Standards and Regulatory Map

6. Documentation Navigation

7. Summary

The four directions of the Embodied Intelligence Group form a complete top-down chain: the brain decides "what to do" (L1/L3), the cerebellum decides "how to do it in real time" (L2/L3 hard real-time loop), the actuators decide "what is physically achievable or not" (L2 end), and materials & sensors decide "what can be seen and touched" (L1 source / L5 entry). Teleoperation data (L4), simulation evaluation (L5), and safety regulations (L6) serve as three lateral supports that stitch the four directions together into an engineering-deliverable physical agent runtime.

The most important industry fact for this group is the "brain/cerebellum separation architecture": Helix's 7–9 Hz / 200 Hz dual system and GR00T N1's dual-system design compress the "slow planning + fast execution" of the Harness Loop from the second-level timing of the software domain to the millisecond-level timing of the physical domain; meanwhile, Galaxy Astral Brain's "unification" route suggests this consensus is not the end point. Whatever direction the route evolves in, this group's Harness bottom line remains unchanged — errors in the physical world are irreversible, and therefore simulation first and real-machine human-in-the-loop are hard constraints, not optional configurations.

8. References

1. 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
2. Report: China Issues the Humanoid Robot and Embodied Intelligence Standards System (2026 Edition) — People's Daily Overseas Edition, 2026-02-28. https://peoplesdaily.pdnews.cn/china/er/30051524844
3. 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
4. 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/
5. Physical Intelligence Official Website (π0 / π0.5 / π0.7 model line) — Physical Intelligence. https://www.pi.website
6. VLA Models Comparison Guide — roboticscenter.ai, 2026. https://www.roboticscenter.ai/guides/vla-models-comparison
7. Report on Figure's Series C Financing — Cailianshe, 2025-09-16. https://m.cls.cn/detail/2147044
8. Figure Official Series C Announcement — Figure AI, 2025-09. https://www.figure.ai/news/series-c
9. Embodied Intelligence: Decoding the Breakthrough Path of China's Robot Industry — Zhonghong Web, 2026-06. https://www.zhonghongwang.com/show-278-464160-1.html
10. Embodied Intelligence Accelerates in Depth — Digital China Web, 2025-12. https://www.digitalchina.gov.cn/2025/xwzx/szkx/202512/t20251230_5263761.htm
11. 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/
12. Unitree Official Support Page (H1 / G1 / M107 motor specifications) — Unitree. https://support.unitree.com/home
13. China Embodied Intelligence Development Report 2025 — East Money PDF, 2026-02. https://pdf.dfcfw.com/pdf/H3_AP202602011819006121_1.pdf
14. LIBERO Benchmark Cross-Comparison (arXiv 2602.18224) — arXiv, 2026-02. https://arxiv.org/pdf/2602.18224
15. Domestic Well-Known Manufacturers of Six-Axis Force Sensors for Industrial Robots — China Electronic Components Industry Association, Sensitive Components and Sensors Branch, 2026-03-10. http://sensor.ic-ceca.org.cn/hangyezixun/508.html