# The Overlooked Non-Consensus Is Whole-Body Unified Control

The overlooked non-consensus is whole-body unified control

Full article: https://haiguangboy.com/posts/gemini-robotics-2

## Core Method

Previous models only controlled a humanoid's upper body for tabletop tasks; this time, for the first time, the entire humanoid is controlled, from feet to fingertips. The official example is Apollo 2 receiving a command like "put the watering can into the green box on the bottom shelf," then walking to the table, picking up the can, taking a few steps to the shelf, and placing it precisely.

The key isn't that it can walk, but that walking, bending, and reaching are all output by the same model. To retrieve something from deep in a shelf, where your foot lands, how your hip rotates, and how your arm extends are coupled; splitting this into a "walking controller + upper-body manipulation model" creates issues at the seams. What's even more telling: the same checkpoint simultaneously drives three embodiments—Apollo 2 with five-fingered hands, Apollo 2 with grippers, and Franka Duo.

## Key Results

Gemini Robotics 2 is the VLA for doing tasks, ER 2 is the high-level brain handling dialogue and decomposing multi-step tasks, and On-Device 2 runs locally. The upper layer is hierarchical: developers can declare the VLA and navigation API side-by-side as agent tools. Currently, only ER 2 is public; the other two are limited to early partners.

## Related Route Comparisons

- Opposite route · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: Architecture: releases three models at once, layered rather than end-to-end integrated contradicts route bet: behavior/action parts must be a unified model, opposing modular assembly
- Same route · [wx_界面新闻_20260605](https://haiguangboy.com/posts/wx_界面新闻_20260605) `wx_界面新闻_20260605_2026_06`: The whole piece has almost no verifiable numbers: success rates exist only in charts, unverifiable at the text level validates boundary: robot task completion rates/independent contribution shares not disclosed; currently more like a data expedition than commercial deployment
- Same route · [苏度科技 WAIC debut: Reality is the endpoint for robots, not the boundary of training](https://haiguangboy.com/posts/sudu-third-position) `wx_晚点latepost_20260718_2026_07`: The whole piece has almost no verifiable numbers: success rates exist only in charts, unverifiable at the text level validates boundary: single founder interview; core tech and commercial claims are self-reported; 20B valuation/10K GPU compute/POC progress have no third-party verification
- Same route · [Predictive features cannot be directly fed to diffusion models](https://haiguangboy.com/posts/leapbot-wa) `leapbot_wa_world_anchor_action_models_via_predictive_latent_alignments_2026_07`: The whole piece has almost no verifiable numbers: success rates exist only in charts, unverifiable at the text level validates real-robot validation is thin: single UR5, only pick-and-place of fruits/vegetables, no success rate numbers or baseline comparisons
- Opposite route · [千寻智能韩峰涛：Embodied intelligence enters the "speed competition" phase](https://haiguangboy.com/posts/qianxun-general-brain-data-loop) `wx_中国财富_20260719_2026_07`: Cross-embodiment costs are publicly priced: a few hours, usually fewer than 200 examples contradicts ★two technical judgments: collection devices must be 'design for AI' and co-developed with the robot body; all resources focused on the data loop
- Same route · [Gemini Robotics ER 2](https://haiguangboy.com/posts/deepmind_blog_gemini_robotics_er_2_20260731) `deepmind_blog_gemini_robotics_er_2_20260731_2026_07`: Access is tiered: only ER 2 is public, VLA and on-device models limited to early partners validates product side: ER 2 goes public preview, capability list includes calling Search and Calendar

## Boundary

Success rates are all in charts; the text only has qualitative descriptions like "medium-to-high success rates," without naming competitors or providing baselines. The official team itself admits in figure captions that dexterous multi-finger manipulation remains difficult and movement speed still needs improvement. The safety report is offline evaluation, with constraint datasets synthesized by the model; real robots were only tested for safe stopping.

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

The debate over "layered or unified" may have been asking the wrong question all along. One camp in the industry argues that behavior/action must be a unified model, opposing modular assembly; another splits physical intelligence into upper and lower layers. Google's approach takes both: the upper layer is split—reasoning and execution are two models connected via tool calls; the lower layer is welded shut—legs and hands are not separable.

So the question isn't whether to layer, but where to make the cut. Cutting between semantics and physics benefits both sides; cutting within physics—say, assigning walking and manipulation to two models—leaves seams exactly where coordination is most needed.

Incidentally, this judgment didn't emerge only after Google's release. Domestic teams had already written "humanoid manipulation cannot be understood as a robotic arm plus a mobile base" into a full technical route, more than half a month before this release. In my next post, I'll break down that one.

Don't rush to conclusions on the cross-embodiment point either: the official team only gave numbers for dual arms in a few hours; whether whole-body humanoids are similar is unstated; other work has also found that raw trajectories struggle to express contact patterns, constraint geometry, and force directions, making transfer at the physical execution layer far less trivial.