Brain

1. Introduction

1.1. Background

The Guiding Opinions on Innovation and Development of Humanoid Robots (MIIT, 2023-11) defines the "brain" as the capability cluster of environment perception, behavior control, and human-robot interaction based on large models, and proposes a technical route of cloud-edge-device collaborative intelligent deployment. On the industry side, the brain is the link with the highest value and funding density in embodied intelligence: Physical Intelligence started with a 2024-03 seed round of $70M and reached a post-money valuation of $5.6B by its 2025-11 Series B (Bloomberg basis); Figure completed a Series C of over $1B in 2025-09 with a post-money valuation of $39B, with the funds explicitly earmarked for Helix expansion, next-generation GPU training infrastructure, and human video + multimodal sensory data collection.

On the brain side, the technology mainline has clearly shifted over the past three years: from the loose coupling of "vision-language model + low-level control," to the end-to-end VLA (Vision-Language-Action) paradigm of "vision-language-action," and then to the dual-system architecture of "slow planning + fast control." Industry reviews summarize it as: the brain decomposes natural-language instructions into logical steps (generalization), while the cerebellum outputs motor control signals at millisecond-level frequencies on local computing platforms (real-time, disturbance-resistant) — this division of labor is precisely the semantic-side expression of this group's "large/small brain separation architecture" assertion.

1.2. Definition and Scope

The brain direction covers the robot's semantic and cognitive layers, and in the AI Harness context specifically refers to the following links in which foundation models participate or lead:

Boundary note: this direction covers only the semantic and planning layers; the 50–200 Hz real-time control loop belongs to Cerebellum (Motion Control). The System 2 part of dual-system VLAs (e.g., Helix, GR00T N1) belongs to this direction, while the System 1 part is discussed in the cerebellum direction.

1.3. Position in the AI Harness System

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

Core judgment: the bottleneck of the brain direction lies in the L5 Evaluation and Observation layer. LIBERO averages of 94%~97% (π0 94.2, π0.5 96.9, OpenVLA-OFT 97.1) describe performance on tabletop manipulation benchmarks; the fact that π0.5 itself makes "completing cleanup tasks in real households never seen during training" its core selling point shows that open-world generalization still lacks a unified, assessable acceptance standard. Another gap lies at the boundary between L1 and L2 — VLA on-device inference latency generally lacks official vendor numbers (only two bases are citable: Helix's 200 Hz control frequency and π0 flow-matching's ~50 Hz), so "how fast the brain can respond" cannot be reconciled on most platforms.

1.4. Current Development Status

The mainline VLA and brain-model lineage is as follows (scores are all paper or official bases, information current to 2026-09-12):

Figure 1-1 | VLA model lineage timeline (2023—2026)

Data sources: Physical Intelligence official site and the arXiv 2607.06706 / 2602.18224 summary tables, NVIDIA official press releases, Figure official and Cailianshe reports; information current to 2026-09-12.

On the world-model side, three main lines are equally clear: NVIDIA Cosmos (announced at CES 2025-01, open-sourced 1.0 weights 1-07, split into Predict / Transfer / Reason series in 2025-03) positions itself as a "world foundation model platform"; Cosmos Transfer converts Omniverse 3D simulation into controllable photorealistic video for synthetic data; Google Genie 3 (2025-08) is the first world model supporting real-time interaction (24 fps, 720p, visual memory traceable back ~1 minute); and World Labs' Marble (commercialized 2025-11) generates fixed 3D scenes using 3D Gaussian splatting. The counter-evidence that must be faced: Physics-IQ measurements show a huge gap between visual realism (best 55.6%) and physical understanding (best 24.1%) — a world model is not equivalent to world understanding.

2. Glossary

3. Case Studies

3.1. Physical Intelligence: Engineering of π0 / π0.5 and General VLA Policies

Background. Physical Intelligence (PI) was founded around 2024, positioned as a "heavy-software brain camp": it does not build embodiments, but focuses on general robot policies. Its funding line moved within two years from a $70M seed round (2024-03, led by Thrive, valuation ~$400M) to a $400M Series A (2024-11, with Bezos, OpenAI Funds, and others participating; two coexisting valuations of $2–2.4B), then to a $600M Series B (2025-11, led by CapitalG, post-money $5.6B, Bloomberg basis); in 2026-03 there were reports of talks for ~$1B in funding at a valuation above $11B — that item is an unconfirmed rumor, included for timeline reference only.

Approach. π0 (2024-10-31) uses PaliGemma 3B + a flow-matching action head to generate smooth continuous action trajectories at ~50 Hz (industry-review basis), completing folding laundry, clearing tabletops, and other tasks across 8 robot configurations; π0-FAST (2025-01) uses a novel action tokenizer to speed up training 5x; Hi Robot (2025-02) introduces hierarchical control, with a high-level model parsing intent and a low-level policy executing actions; π0.5 (2025-04-22) focuses on open-world generalization. The model line continues to π*0.6 (2025-11, a VLA that learns from experience via reinforcement learning) and π0.7 (2026-04, compositional generalization: learns to operate an air fryer from 2 examples).

Results. On the LIBERO benchmark, π0 averages 94.2 (two-literature aggregate basis: Spatial 96.2/96.8, Object 97.1/98.8, Goal 93.5/95.8, Long 78.4/85.2) and π0.5 averages 96.9; flow-matching class models show notable advantages on long-horizon tasks (LIBERO-Long) (arXiv 2602.18224 Table 2 basis). π0 / π0-FAST weights and code were open-sourced on 2025-02-04, becoming one of the de facto bases for industry reproduction and fine-tuning. Confidence: model scores are of paper and official basis (high); funding valuations are of financial-media basis (medium-high).

3.2. Figure Helix: Dual-System Architecture and a $39B Valuation

Background. Figure is a representative of the "full-stack, software-and-hardware" camp. In 2025-02 it ended its partnership with OpenAI and turned to in-house development; that same month it released Helix — the first publicly known "dual-system" VLA. On 2025-09-16 it completed a Series C of over $1B with a post-money valuation of $39B (led by Parkway, with NVIDIA, Intel Capital, LG, Salesforce, T-Mobile, Qualcomm, Brookfield, Macquarie, and others participating); its earlier Series B (2024-02) was $675M at a valuation of ~$2.6B.

Approach. Helix consists of two systems: the slow system (System 2) is an onboard pretrained VLM running at 7–9 Hz, responsible for scene understanding and language understanding; the fast system (System 1) outputs complete control of the robot's upper body at 200 Hz (wrist pose, finger flexion, torso/head orientation toward targets). This "frequency-layered" design decouples semantics from control — the semantic layer can be slow, but the control layer must be fast. The parameter count is not disclosed. Series C funds are earmarked for Helix expansion, next-generation GPU training infrastructure, and human video and multimodal sensory data collection, pointing upstream to data and compute rather than a single model.

Results. Helix's 7–9 Hz / 200 Hz dual-system frequencies have become the most-cited public parameters of the "large/small brain separation architecture" (official basis); the capital market has given it the industry's highest valuation anchor. It should be noted that Helix is closed-source with no publicly comparable benchmark scores; its capability evidence comes mainly from official demo videos, and third-party reproduction data is missing — citations should be labeled "official demo basis." Confidence: funding and architecture parameters are of official and Cailianshe basis (high); capability claims have medium confidence.

3.3. Galbot S1: The Brain of the Simulation Synthetic-Data Route

Background. Galbot is the domestic representative of the "heavy-software brain camp," taking the opposite route from schools that depend on real-robot teleoperation data: it centers on 3D visual synthetic data + Sim2Real training. In 2025-12 it completed a new financing of over $300M, cumulatively ~$800M, at a valuation of $3B (Baidu Baike basis); afterward it raised another RMB 2.5B round (with the National AI Industry Investment Fund and others; Huachuang Securities report basis, early 2026, claiming cumulative financing ranks first among Chinese embodied-intelligence companies). The two financing bases are listed in parallel by timeline and not merged.

Approach. It has self-developed "Galaxy Star Brain" (AstraBrain), claiming to be "the world's first full-body full-hand end-to-end embodied large model integrating brain, cerebellum, and neural control into a single model" — this is the reverse route of the mainstream "large/small brain separation" assertion, showing that whether the large and small brains are separated remains an open question in 2025—2026. The technical base is grasping and manipulation policies trained on pure simulation synthetic data, paired with the Open6DOR simulation platform (2,500-task detection); per the Baike basis, its 6-DoF manipulation precision error is less than 0.5 mm.

Results. Per company and Baike basis (lacking third-party reproduction, medium confidence): Galbot S1 achieves fully autonomous routine operation at CATL's battery factory; transparent-object grasping success rate is 95%; at the 2025 Beijing BAAI Conference it performed a continuous 18-hour real-robot demo with a 97% task success rate. On the commercialization side, as of 2025-06 it covered 7 unmanned pharmacies in Beijing with over 300 daily orders; in 2025-08 it won the pharmacy-sorting competition of the World Humanoid Robot Games (4 min 25 s in the preliminary round). This case shows that the synthetic-data route can work in structured commercial scenarios (pharmacies, factories), but its 95% / 97% figures are all company basis and must be labeled when cited.

4. Practical Standards

4.1. AGENTS.md Specification (Brain Direction)

The following is the complete copyable AGENTS.md content for the brain direction, a trimmed and strengthened version of the group-level AGENTS.md:

# AGENTS.md —— 具身智能组 · 大脑方向

## 角色与边界
- **角色**：大脑方向工程智能体，负责 VLA 与世界模型的训练、评测、部署与任务规划链路开发。
- **边界**：不直接输出电机控制信号（50–200 Hz 实时环归小脑方向）；真机部署只产出待人工确认的指令包。
- **第一原则**：基准成绩不等于真机能力。任何能力声明必须基准成绩与真机验收双报告。

## 环境假设
- 声明基座模型与版本（π0 / OpenVLA / GR00T N1 等）、微调配置、评测基准版本（LIBERO 四套件 / HumanoidBench 27 任务）。
- 声明部署端算力（如 Jetson Orin / 机载 x86）与推理频率预算；无官方时延数字时标注 [待填写]，不得估算冒充。
- 声明数据来源（Open X-Embodiment / RT-X / AgiBot World / 自采）与合规状态。

## 上下文加载顺序（Context Budget）
- 必载：任务判据、模型配置摘要、回归集版本与当前最好成绩。
- 次载：数据集元信息（规模、构型、采集方式）、失败案例清单。
- 禁止：原始轨迹数据、完整模型权重张量进入上下文。

## 工具契约
- 训练与评测脚本须版本锁定、随机种子固定、结果可复现。
- 评测结果必须含退出码、日志路径、分套件成绩（Spatial / Object / Goal / Long 分列）。
- 公司自报成绩不得作为本项目评测结果引用，只能作为外部对照并标注"公司口径"。

## 任务执行流程（SOP）
- S1 解析任务为可判定子目标；S2 固定环境复现基线；S3 微调或评测；S4 分套件报告并与基线对照；S5 真机部署走组级红线的影子验证与人在回路流程。

## 验证与证据要求
- 双报告制：基准成绩 + 真机验收（缺真机数据须显式声明"仅仿真/仅基准"）。
- 长程任务单列 LIBERO-Long 等长程成绩，禁止只报四套件均值掩盖长程短板。
- 融资、估值等商业数字标注媒体口径与两口径差异。

## 失败与升级策略
- 微调后成绩低于基线：检查数据质量与格式，冻结训练做最小复现。
- 真机行为与仿真不一致：回到组级红线，走 sim2real 排查，不绕过小脑限幅。

## 安全与合规红线
- 输出动作须经安全语义过滤：拒绝危险指令、限制动作语义边界。
- 训练数据涉人脸、家庭场景的须脱敏；对外发布前过合规审查。

## 禁止事项
- 禁止把官方演示视频描述为"实测成绩"；禁止混用仿真与真机数字。

## 输出格式
- 模型卡格式：架构、数据、基准分套件成绩、部署约束、已知失败模式、开源状态。

## 评估与自检
- 自检项：双报告、口径标注、种子可复现、长程成绩单列、数据合规、无占位符。

4.2. SKILL.md Specification (Brain Direction)

---
name: embodied-brain-vla
description: 大脑方向技能。当需要选型、微调或评测 VLA 模型，搭建任务规划链路，或生成世界模型合成数据，并要求基准与真机双报告时使用。
version: 1.0
created: 2026-09-12
---

# 大脑方向技能

## 适用场景
- VLA 模型选型与对比（RT-2 / OpenVLA / π0 系列 / GR00T N1 / Helix）。
- 少样本微调（OpenVLA 类：100–200 条示教）与评测。
- 任务规划链路（分层控制）与世界模型合成数据生成（Cosmos Transfer 等）。

## 前置条件
- 评测基准与脚本版本固定；部署算力与频率预算明确。
- 数据合规状态可追溯。

## 输入
- 任务目标与成功判据；基座模型与配置；数据集路径与版本。

## 输出
- 模型卡 + 分套件评测报告 + 可复现配置（种子、版本、脚本路径）。

## 执行步骤
1. 固定基线：复现基座模型在回归集上的成绩。
2. 微调或评测：多种子执行，分套件统计（含长程套件）。
3. 对照外部口径（公司自报成绩仅作参考并标注）。
4. 部署前走影子验证；真机流程移交组级红线流程。
5. 产出模型卡与遗留风险清单。

## 质量标准（DoD）
- 双报告制；种子可复现；长程成绩单列；口径全部标注；无占位符。

## 常见失败与处理
- 微调退化：先查数据格式与质量，再做最小复现。
- 基准高分、真机失败：按 sim2real 排查，禁止只报基准。

## 示例
- 用 100 条自采数据微调 OpenVLA，在 LIBERO-Spatial 上评测并与开源基线 76.5% 对照。

4.3. Implementation Checklist

5. Summary

The brain direction is the link where embodied intelligence and AI Harness mesh most tightly: VLM perception assembly corresponds to L1, task decomposition and skill scheduling to L3, embodied memory to L4, benchmarks and open-world acceptance to L5, and semantic safety filtering to L6 — the brain itself is the slow system of a Harness Loop running in the physical world.

The engineering conclusions from the three-year lineage are threefold. First, end-to-end VLA has become the default paradigm: from RT-2's action Tokenization to π0's flow matching, then to GR00T N1's open base model, action generation has completed the shift from "semantically describing actions" to "directly generating continuous actions." Second, the dual-system architecture has become the common choice of both the open and closed camps: Helix's 7–9 Hz / 200 Hz and GR00T N1's System 2 / System 1 are structurally isomorphic, validating the feasibility of the "slow semantics, fast control" layering; the Galaxy Star Brain's unified route is the counter-evidence that must be continuously tracked. Third, evaluation is currently the weakest layer: the gap between LIBERO averages of 94%~97% and real open-world success rates, and the general absence of official on-device inference-latency numbers, both require the brain-direction Harness to treat "benchmark scores + real-robot acceptance + measured frequency" as the minimum delivery standard.

Information-Gap Statement

The following items have not been confirmed by primary sources and have been labeled in the main text:

1. Official vendor numbers for VLA on-device inference latency: only Helix's 200 Hz (control frequency) and π0's ~50 Hz (industry-review basis) are citable; OpenVLA and others have no official first-hand latency data on on-device platforms such as Jetson Orin, [To be filled].
2. The ~$1B funding and above-$11B valuation rumored for Physical Intelligence in 2026-03: reported but unconfirmed; timeline reference only.
3. Helix's parameter count and public benchmark scores: not disclosed officially; capability evidence is official-demo basis only, medium confidence.
4. Galbot's 95% transparent-object grasping, 18-hour 97% success rate, and 6-DoF manipulation error under 0.5 mm: company/Baike basis, lacking third-party reproduction, medium confidence.
5. π0's per-suite LIBERO scores differ slightly between the two summary tables in arXiv 2607.06706 and 2602.18224; this document lists them in parallel without choosing one.
6. The original URL of Unitree's prospectus was not obtained; the brain-related full-stack wording ("hardware + cerebellum + brain") comes from a China Economic Net report summary.

6. References

1. Physical Intelligence official site (the π0 / π0-FAST / π0.5 / π*0.6 / π0.7 model line) — Physical Intelligence. https://www.pi.website
2. VLA Models Comparison Guide — roboticscenter.ai, 2026. https://www.roboticscenter.ai/guides/vla-models-comparison
3. LIBERO benchmark cross-comparison table (arXiv 2602.18224) — arXiv, 2026-02. https://arxiv.org/pdf/2602.18224
4. π0-series LIBERO score summary (arXiv 2607.06706) — arXiv, 2026-07. https://arxiv.org/html/2607.06706v1
5. OpenVLA-OFT paper page — 2025. https://kendrick-stein.github.io/MCISLAB_DeepRead/Papers/2502-OpenVLA-OFT
6. 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
7. 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/
8. Figure official Series C announcement — Figure AI, 2025-09. https://www.figure.ai/news/series-c
9. Figure Series C financing report — Cailianshe, 2025-09-16. https://m.cls.cn/detail/2147044
10. Guiding Opinions 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
11. Physical Intelligence financing and model-line chronology — The Dynamics, 2026. https://thedynamics.ai/articles/physical-intelligence-history
12. Robotera official milestones (brain technology line) — Robotera. https://www.robotera.com/#/about/us
13. Embodied intelligence: decoding the path to breakthroughs in China's robot industry — Zhonghong Net, 2026-06. https://www.zhonghongwang.com/show-278-464160-1.html
14. 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
15. ALOHA 2 / Mobile ALOHA entries — aiwiki, 2024–2025. https://aiwiki.ai/wiki/aloha_2