GigaBrain-WBC-0.5 A Behavior World Model for Robust Whole-Body Control with Environment Interaction

Ziyang Cheng^1,2 Tianshu Tang^1,2 Jinxin Lan^1,2 Xinze Chen² Yuhan Gong² Zhichao Liu² Changzhong Wu² Yahao Mao^2,3 Zongyan Deng^2,4 Mingxuan Ma^2,4 Huasen Xi^2,5,6 Yilong Liu^1,2 Yutong Wu² Xiaofeng Wang² Yang Wang² Yun Ye² Guan Huang² Xiaojie Jin⁴ Zheng Zhu^2,# Jiwen Lu^1,#

¹Tsinghua University ²GigaAI ³University of Shanghai for Science and Technology ⁴Beijing Jiaotong University ⁵Institute of Automation, Chinese Academy of Sciences ⁶University of Chinese Academy of Sciences ^#Corresponding authors

Paper Codecoming soon Videos

Full system

One policy. Real-time command. Real-world contact.

Full demo Real-robot footage forthcoming

GigaBrain-WBC-0.5 models what the robot can do next while it acts, enabling environment interaction and extreme robustness in a single whole-body controller.

Head-to-head

Comparison with SONIC

Matched hardware trials: SONIC is shown on the left and ours on the right, driven simultaneously by the same live command.

Box liftingObject interaction

Lifting a box

Ours lifts and holds the box while SONIC loses a stable grasp.

Fire extinguisherObject interaction

Kneel-to-stand with payload

Ours rises while holding a 2 kg fire extinguisher; SONIC falls during the ascent.

Step onto platformTerrain interaction

Stepping onto a platform

Ours climbs onto the platform and stabilizes; SONIC fails to complete the ascent.

Sit on a boxTerrain interaction

Sitting on a support

Ours loads the box as a seat; SONIC remains in a half-squat without stable support.

Capability 01

Environment interaction

The controller turns seats and platforms into load-bearing support while coordinating balance with carried objects.

Chair supportLoad-bearing contact

Sitting on a chair

The robot lowers onto a real chair and uses the seat as load-bearing support under live whole-body command.

Case + platformObject and terrain interaction

Carrying a case onto a platform

While holding a large black case, the robot steps onto the platform and stabilizes without dropping it.

Box + platformPayload and terrain interaction

Carrying a box onto a platform

The robot climbs while carrying a cardboard box, adapting its balance to both the payload and the platform contact.

Capability 02

Extreme robustness

When the command under current environment is infeasible, the policy remains driveable.

Chair removedBest-effort squat

Missing support

The same sitting command resolves to a stable squat.

Step removedContinuous task

Changed terrain

The robot keeps carrying instead of committing to a missing step.

DisturbancePush, trip, recover

Physical disturbance

Push and trip recovery within the tracking policy itself.

OOD commandClosest feasible behavior

Unsafe commands

Infeasible intent is retracted to a continuous best-effort motion.

Motion range

High-difficulty motion

One command interface across flexibility, control, and explosive dynamics.

High kick

High kick

Tai chiSlow control

Tai chi

JumpDynamic take-off

Jumping

Embodiment

Cross-robot transfer

The same controller design transfers beyond a single humanoid body.

Unitree G1Source embodiment

Unitree G1

Maker L01Transferred embodiment

Maker L01

Positioning

One policy closes the capability gap

Prior methods cover pieces of the problem. GigaBrain-WBC-0.5 brings them together.

Method	Diverse tracking	Teleoperation	Terrain interaction	Object interaction	OOD robust	Fall robust
GMT	✓	✕	✕	✕	✕	✕
TWIST	✓	✓	✕	✕	✕	✕
SONIC	✓	✓	✕	✕	✕	✕
HoloMotion-1	✓	✓	✕	✕	✕	✕
Humanoid-GPT	✓	✓	✕	✓	✕	✕
SceneBot	✓	✕	✓	✓	✕	✕
CMP	✓	✓	✕	✓	✓	✕
BFM-Zero	✓	✓	✕	✕	✕	✓
GigaBrain-WBC-0.5	✓	✓	✓	✓	✓	✓

Core results

Strong across every regime

Best values are highlighted. All results are sim-to-sim in MuJoCo.

4.3×terrain survival over the strongest baseline

83.1%survival under physically implausible commands

99.3%recovery from fallen initializations

Method	Standard			Terrain			OOD		Fall
Method	MPKPE↓	RootVel↓	SR↑	MPKPE↓	RootPos↓	SR↑	MPKPE↓	SR↑	SR↑	Jerk↓
SONIC	82.3	189.6	94.1	331.2	294.7	15.3	327.6	50.0	5.9	1295.5
HoloMotion-1	109.4	121.3	89.0	330.0	280.7	18.7	248.7	67.7	0.7	2000.0
Humanoid-GPT	90.9	205.8	91.9	283.3	326.7	14.0	208.0	70.6	2.9	3598.1
Ours	76.6	211.1	96.3	93.3	100.7	81.3	158.0	83.1	99.3	1050.6

Errors are in mm except RootVel (mm/s) and Jerk (rad/s³); SR is percent. Fall SR uses the recovery criterion described in the paper.

Method overview

A controller that models its own future behavior

A causal Transformer jointly predicts action, next proprioceptive state, and the distribution of the next latent behavior command. The same prediction used to act also defines what is feasible now.

GigaBrain-WBC-0.5 behavior world model overview — The policy internalizes contact dynamics and behavior admissibility, then reuses its predicted distribution for online best-effort command projection.

Data pipeline

From motion to spatial terrain

Contact evidence in retargeted motion is recovered as full 3D, simulator-ready geometry: stairs, seats, boxes, tables, and other supports.

Automatic spatial terrain annotation pipeline — Contact detection, whole-body penetration filtering, clustering, and primitive fitting turn ordinary motion data into a terrain-paired training corpus.

Abstract

Whole-body motion tracking policies turn a humanoid into a robust control interface: the teleoperator—or an upstream model—only supplies a coarse movement intent, while the low-level policy keeps the robot balanced and physically feasible.

Existing trackers deliver this interface only on flat ground: trained in empty scenes, they never learn how contact with terrain and objects reshapes their dynamics, and they attempt to teach the policy to balance under any command by continually enlarging the reference-motion corpus, which stops working once feasible behaviors become environment-dependent.

We present GigaBrain-WBC-0.5, the first Behavior World Model (BWM) for humanoid whole-body control. Rather than a purely reactive tracker, we train a causal Transformer to jointly predict its next action, next state, and the distribution over its next latent behavior command, so the network that acts also models how the environment shapes what it can do next.

An automatic terrain-annotation pipeline recovers full 3D contact geometry from retargeted motion, enabling terrain annotation at the scale of existing motion datasets.

The predicted distribution is reused at deployment to detect implausible commands online and retract them onto learned behaviors, so the robot attempts tasks in a “best-effort” manner.

The result is a unified policy that takes real-time command, interacts with environment, and stays robust to implausible commands, falls, and disturbances.

GigaBrain-WBC-0.5 achieves the highest success rate across all four regimes among three large-scale tracker baselines: 81.3% on terrain interaction (4.3× the strongest baseline), 83.1% under implausible commands, and 99.3% fall recovery (16.8× the strongest baseline). Hardware trials show robust interaction under missing supports and disturbances; the Unitree G1 checkpoint transfers to the Maker L01 robot with simple fine-tuning.

Citation

BibTeX

@article{gigabrainwbc2026,
  title   = {GigaBrain-WBC-0.5: A Behavior World Model for Robust
             Whole-Body Control with Environment Interaction},
  author  = {Cheng, Ziyang and Tang, Tianshu and Lan, Jinxin and Chen, Xinze and Gong, Yuhan and Liu, Zhichao and Wu, Changzhong and Mao, Yahao and Deng, Zongyan and Ma, Mingxuan and Xi, Huasen and Liu, Yilong and Wu, Yutong and Wang, Xiaofeng and Wang, Yang and Ye, Yun and Huang, Guan and Jin, Xiaojie and Zhu, Zheng and Lu, Jiwen},
  journal = {Technical Report},
  year    = {2026}
}