Brace Yourself: Task-Conditioned Environmental Bracing for Forceful Humanoid Manipulation

Zongyuan Zhang, Christopher Lehnert, Will N. Browne, Jonathan M. Roberts · N/A · 2026

Summary

Abstract Summary

Forceful manipulation is challenging for humanoid robots because interaction forces can disturb whole-body balance. We introduce the Supporting Hand Strategy (SHS), which enables a humanoid to brace against the environment with one hand while performing forceful manipulation with the other. SHS optimises a task-conditioned support configuration that guides two synchronous reinforcement-learning policies, without human motion data or online whole-body trajectory planning. On a Unitree G1, SHS achieved usable contact forces up to 60 N, compared with a maximum sustained force of 13.5 N without environmental bracing, while substantially improving force tracking over a task-independent support configuration. The same policies generalised to different task regions without retraining. SHS therefore provides a simple mechanism for substantially extending humanoid forceful-manipulation capability.

Key Points

Forceful manipulation is challenging for humanoid robots because interaction forces can disturb whole-body balance.
We introduce the Supporting Hand Strategy (SHS), which enables a humanoid to brace against the environment with one hand while performing forceful manipulation with the other.
SHS optimises a task-conditioned support configuration that guides two synchronous reinforcement-learning policies, without human motion data or online whole-body trajectory planning.
On a Unitree G1, SHS achieved usable contact forces up to 60 N, compared with a maximum sustained force of 13.5 N without environmental bracing, while substantially improving force tracking over a task-independent support configuration.
The same policies generalised to different task regions without retraining

Supporting Hand Strategy

SHS separates forceful manipulation into offline support-configuration optimisation and RL policy execution. It uses three components: the Task-Region Support Optimiser (TRSO), the Whole-Body Keypoint Representation (WBKR), and synchronous body and working-arm policies.

TRSO finds one support configuration that accommodates the working-hand position–force requirements across a task region while placing the supporting hand within an admissible support-contact region. WBKR encodes this configuration geometrically for policy input.

Supporting Hand Strategy — Brace Yourself: Task-Conditioned Environmental Bracing for Forceful Humanoid Manipulation — Figure 5: Matched hardware trajectory tracking in T1 at 55 N: (a) SHS and (b) Generic. Black curves show the target trajectory; blue/orange curves show executed laps, with marker colour denoting mean measured normal force. At this load, SHS follows the target trajectory more closely than Generic.

Experimental Design

The ablations isolate three components of SHS: task-conditioned support configuration, WBKR guidance, and environmental bracing. Table I summarises the comparisons. The four RL methods share the same dual-policy architecture, task distribution, dynamics, PPO settings, and training budget; ablated inputs and reward terms differ only as described below.

Results

Across 30 paired simulation runs, SHS outperformed No-bracing in every pair: environmental bracing raised episode success rate from $0.020$ to $0.666$ and mean stable-contact hold ratio from $0.069$ to $0.950$ (Fig. 3 (a,b)). In Experiment 1, SHS reached usable forces of 60 N at the centre and upper positions and 40 N at the lower positions, compared with 40–50 N for Generic (Fig. 4 (a)).

Modular, the unsupported hardware baseline, was usable only at 10 N at the lower-right position and never sustained more than 13.5 N for one second.

Sources and Demonstrations

Figures are reproduced from Zhang et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.