1
2
3
When a humanoid robot can no longer keep its balance, the fall itself becomes the problem. Inspired by martial-arts breakfalls (ukemi), Ukemi-SafeFall teaches the robot to roll through the impact and then stand back up.
Simulation, fallen-state protocol (N = 50). The first three are reductions vs. Passive Fall.
Humanoid robots fall when balance recovery fails, and most controllers treat that as the end of the task. We handle it instead: Ukemi-SafeFall learns to reduce the mechanical damage of an unavoidable fall while keeping the robot able to stand up afterward. In simulation it cuts normalized peak contact force by 32%, cumulative peak contact force by 78% and unsafe-contact events by 70% compared with passive falling, and it transfers to a physical humanoid across a range of falling scenarios.
Rolling and recovery rewards switch on automatically from torso tilt and head height, so no separate phase estimator is needed.
Training penalizes heavy loading on the trunk and arms, which lowers peak contact force and the number of unsafe contacts.
A frozen Score-Matching Motion Prior scales the reward, keeping the learned motion close to natural human behavior.
We train a policy with Proximal Policy Optimization. Its reward multiplies condition-gated task rewards (safe rolling, impact mitigation, standing up) by a reward from a diffusion-based Score-Matching Motion Prior. The same prior also generates diverse fallen starting poses for training.
All methods are evaluated on the same fallen-state protocol. We compare against Passive Fall, remove one component at a time in three ablations, and include a re-implemented SafeFall baseline as a reference.
Table. Ablation and comparison methods.
| Method | Ukemi task reward | Impact-mitigation | Score-Matching Motion Prior reward |
|---|---|---|---|
| Passive Fall | × | × | × |
| Ukemi-SafeFall: Recovery Only | × | × | ✓ |
| Ukemi-SafeFall without Condition Gating | Always active | ✓ | ✓ |
| Ukemi-SafeFall without Score-Matching Motion Prior Reward Gating | ✓ | ✓ | × |
| SafeFall Triangle Proximal Policy Optimization (*) | × | ✓ | × |
| Ukemi-SafeFall (Ours) | ✓ | ✓ | ✓ |
(*) Evaluated and reported as a cross-protocol reference
The controller outputs zero action, showing what an uncontrolled fall looks like.
Trained only to get up, without rolling or impact-mitigation rewards.
The ukemi rewards stay active for the whole episode instead of only when needed.
Keeps the gated task rewards but removes the motion-prior reward gate.
A protective-fall baseline with no recovery objective. It uses a different training protocol, so it is a reference only.
The full method: gated ukemi rewards, contact-force shaping and the motion-prior gate.
We test falls that arise from walking, running, stair descent and platform jumps, first in a separate MuJoCo environment and then on a physical humanoid.
Qualitative demonstrations on our in-house humanoid with 23 actuated degrees of freedom. They show that the behavior transfers to hardware; the quantitative metrics come from simulation.
@inproceedings{tran2026ukemisafefall,
title = {Ukemi-SafeFall: An Ukemi-inspired Injury-Aware Falling and Recovery},
author = {Tran, Duc Anh and Vu, Quang-Anh and Nguyen, Hieu T. and Pham, Loc and Nguyen, Le-Chuong and Dao, Quang Huy and Vu, Dao Hoang and Nguyen, Truong and Nguyen, Chuong and Le, Tuyen P. and Nguyen, Quan},
booktitle = {2026 IEEE-RAS 25th International Conference on Humanoid Robots (Humanoids)},
year = {2026},
note = {Tran Duc Anh and Quang-Anh Vu contributed equally.}
}