---
title: "The Whole-Body Control Race: Beijing's Hi-WBC Joins Nvidia's GR00T 2 and Google's Gemini Robotics 2.5"
date: 2026-11-18
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-hi-wbc-beijing-humanoid
language: en
---

# The Whole-Body Control Race: Beijing's Hi-WBC Joins Nvidia's GR00T 2 and Google's Gemini Robotics 2.5

> Three major whole-body and manipulation control releases landed within nine days — Beijing X-Humanoid's Hi-WBC, Nvidia GR00T 2, and Gemini Robotics 2.5 on a Mercedes line — and each attacks the same problem from a different direction.

A humanoid robot can perform a flawless backflip on stage and then fail at the task that actually pays: carrying a heavy box while someone bumps into it, or stepping over a cable on an uneven floor. That gap between choreography and work is the target of three releases that landed within nine days of each other this November — from a Beijing state-backed lab, from Nvidia, and from Google DeepMind — and together they map the current state of the whole-body control race.

## Beijing's answer: treat motion reference as a prior, not a script

On November 2, the Beijing Innovation Center of Humanoid Robotics (北京人形机器人创新中心, X-Humanoid) released Hi-WBC, a general whole-body control model, through its "Huisi Kaiwu" (慧思开物) embodied AI platform. Its core conceptual move: stop treating human motion capture as a script the robot must copy exactly. Instead, the human reference becomes a motion prior — an intent to preserve — while physics decides the execution. The center's own framing is "rigid and flexible at once": follow the intent precisely, but adapt posture, joints, and support points to load, force, and terrain.

Three technical pillars back that idea:

- **Skill expansion via Mixture-of-Experts and policy distillation.** Specialist expert policies learn specific tasks — running, carrying, climbing over obstacles — and their abilities are distilled back into the general policy, so new skills accumulate without erasing old ones. This attacks the industry's re-tuning problem: adding a task no longer means rebuilding the controller.

- **Contact-aware learning.** The robot estimates external forces and payload changes from proprioceptive signals like joint currents and motion states — no extra force sensors required — combined with short-horizon dynamics optimization, coordinating posture and support in real time under load or shoving.

- **Perceptive control for terrain.** Terrain geometry is embedded directly in the control loop: the robot identifies feasible footholds on slopes, stairs, and gaps, then rebuilds foot placement, swing trajectories, and body posture while keeping the original movement intent.

The evidence offered is practical rather than spectacular. At a robot athletics competition, Tien Kung (天工) robots running Hi-WBC took gold and silver in the "comprehensive five-event" contest, where one robot must chain walking, height adjustment, grasping, and carrying into a single continuous run — and was the only bipedal humanoid to win gold in the scenario track. A separate one-shot demo shows multiple robots dividing up a long-horizon conference-room cleanup: moving chairs, picking up litter, carrying a garbage bag, squatting to collect paper scraps, with handoffs between machines. X-Humanoid also says the first national standards for humanoid motion control, which it led in drafting, have been formally released.

## The international context, eight days apart

Nvidia entered on November 9 at the Conference on Robot Learning in Austin with GR00T 2, an open world action model built on its Cosmos foundation. Where Hi-WBC perfects the body, GR00T 2 attacks the data problem: it predicts future visual frames and actions jointly, pre-trained on roughly 100,000 hours of open human egocentric video plus teleoperation data, so skills can transfer across robot bodies without new demonstrations for each platform. Nvidia reports that on 50 evaluated tasks the model completed 40 at least once with zero task-specific demonstrations; with one demonstration per task, average success reached 71%, rising to 95% with five. Those are Nvidia's own numbers, and the weights and code are slated for release in December — meaning independent verification has not yet happened. It ships alongside Isaac Lab 3.0, the updated open-source simulation framework that adds deformable-object simulation and reported training-throughput gains.

A day later, Google DeepMind released Gemini Robotics 2.5, which it calls its most capable vision-language-action model, evaluated across more than 200 real tasks with reported gains of 30% to 100% in task success over baselines and 20% to 30% in speed. The headline deployment: Apptronik's Apollo 2 humanoid, controlled by Gemini Robotics 2.5 together with the Gemini Robotics ER task-tracking model, autonomously sorting metal brackets from bulk bins into sequence trays on live Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany — a reported 96% task success rate. DeepMind describes this as a proof-of-concept and has not disclosed run durations, trial counts, or line-takt comparisons.

## Three philosophies, one bottleneck

The interesting pattern is what each team chose to optimize. X-Humanoid, which also builds its own robots, deepens the control layer itself — the part that turns motion plans into stable physics under load and terrain. Nvidia optimizes for data efficiency and portability across embodiments, betting that skill reuse is the industry's cost bottleneck. DeepMind leans on scale of a general VLA plus a division of labor between a reasoning model and a control model, and takes it directly into a working auto plant.

## Honest limitations

All three claims share a common weakness: every headline number — Hi-WBC's competition wins, GR00T 2's 95%, the 96% on the Mercedes line — is self-reported by the party selling the technology, and none yet includes the metrics that industrial buyers actually price: hours of continuous operation, failure recovery rates, maintenance load, and return on investment versus dedicated automation. The race is real, but so is the distance between a demonstration and a shift.

## What readers can do now

- **Demand continuity metrics, not success rates.** When evaluating any of these systems, ask for hours-per-failure and recovery behavior; a 96% single-task rate says little about an eight-hour shift.

- **If you work with Chinese humanoid platforms, test the MoE skill path.** Hi-WBC's claim that new skills can be added without re-tuning the whole controller is checkable by adding one task and measuring regression on old ones.

- **Track December's GR00T 2 release.** Whether Nvidia actually ships open weights, and under what license, will determine if cross-embodiment transfer claims can be independently tested — and whether small labs get access.

- **Watch the standards track.** X-Humanoid's leadership in China's first national humanoid motion-control standards is a quiet but consequential signal for anyone exporting into that market.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-hi-wbc-beijing-humanoid
Free to quote with attribution and a link to the original.
