# Teacher Li Hongyang's Whole-Body Intelligence Surpasses GR's Whole-Body Control

Teacher Li Hongyang's whole-body intelligence surpasses GR's whole-body control

Full article: https://haiguangboy.com/posts/archon-whole-body-intelligence

## Core Approach

S2 handles tasks and semantics, converting open-ended instructions into staged goals; S1 is a native humanoid foundation model, ingesting vision, language, proprioception, and touch, outputting whole-body action intent; S0.5 combines motion generation with behavior foundation models, turning intent into executable reference motions; S0 comprises trackers and controllers, managing balance, contact, and force on real hardware

"Native humanoid" is the key term: the model must from the start understand how head observation affects hand manipulation, how foot placement affects reachable space, and how torso posture yields to manipulation

## Key Results

The data division is summarized as: human data provides scale, robot data provides boundaries, and failure data provides recovery. The four dimensions of scaling are modality, body, task, and failure—parameter count is not among them. The author also offers seven self-check criteria for evaluation, with the core being not to look at single demonstrations but at the speed of capability improvement on new scenes, new hardware, and new tasks

## Comparison of Related Approaches

- Opposite approach · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: Definition: whole-body intelligence is a pretraining paradigm, explicitly not equal to whole-body control, nor is it stitching together VLA/BFM/controllers contradicts joint co-training of heterogeneous human-robot data is a structurally suboptimal solution
- Opposite approach · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: Four-layer stack S2/S1/S0.5/S0, each layer with clear interfaces contradicts route bet: the behavior/action part must be a unified model, opposing modular assembly
- Opposite approach · [wx_界面新闻_20260605](https://haiguangboy.com/posts/wx_界面新闻_20260605) `wx_界面新闻_20260605_2026_06`: A middle stance on software-hardware coupling: hardware-aware pretraining, neither pursuing hardware independence nor allowing coupling contradicts route bet: build the brain, not the body—betting on the foundation model company identity, rejecting the motion capability track
- Opposite approach · [The overlooked non-consensus is whole-body unified control](https://haiguangboy.com/posts/gemini-robotics-2) `deepmind_blog_gemini_robotics_2_brings_20260730_2026_07`: S1's key claim: native humanoid is not a robotic arm plus a mobile base, nor a VLM followed by an action head contradicts architecture: issuing three models at once, layered rather than end-to-end integrated
- Same approach · [Gemini_Robotics_2_Safety](https://haiguangboy.com/posts/gemini_robotics_2_safety) `gemini_robotics_2_safety_2026_07`: Division of labor: whole-body control solves execution problems, whole-body intelligence solves body organization problems validates empirical finding: there is a clear gap between the model "understanding constraints" and "acting according to constraints"
- Same approach · [Introducing Gemini Robotics ER 2](https://haiguangboy.com/posts/google_blog_gemini_robotics_er_2_20260730) `google_blog_gemini_robotics_er_2_20260730_2026_07`: Four-layer stack S2/S1/S0.5/S0, each layer with clear interfaces validates architecture description: VLA models and navigation APIs are declared side by side as agent tools

## Boundaries



## Author's Judgment (Not Paper Content, Cross-Paper Synthesized View)

The most interesting part is not who it agrees with, but that it takes the opposite stance on two levels

📚 On the question of "whether to split," it aligns with the faction advocating unified models and opposing modular assembly—both believe humanoid manipulation cannot be cut along module boundaries. Yet the four-layer stack it proposes is precisely a layered architecture, aligning with GR's approach of splitting the upper layers

🧲 These two things are not contradictory because they address different cuts. It opposes cutting functions horizontally like "navigation module plus grasping module," and advocates cutting abstraction levels vertically like "semantics, body priors, motion generation, execution." The former would break coupled actions apart; the latter would not

🎯 So this judgment can be more precise: the disagreement is not about whether to layer, but whether the cut is along functional boundaries or abstraction levels. Cutting along functional boundaries assigns walking and reaching to two models; cutting along abstraction levels keeps whole-body coordination within the same layer

🧬 There is another point worth noting separately: it argues the model must know what hardware it stands on—joint reachable ranges, force control precision, and sensor perspectives must all enter the priors. This is a bet in the opposite direction from "the same model swaps bodies in a few hours," and in half a year we will see who is right!