IEEE-RAS Humanoids 2026 · Accepted

Ukemi-SafeFall: An Ukemi-inspired Injury-Aware Falling and Recovery

Tran Duc Anh1,*, Quang-Anh Vu1,*, Hieu T. Nguyen1, Loc Pham1, Le-Chuong Nguyen1, Quang Huy Dao1, Dao Hoang Vu1, Truong Nguyen2, Chuong Nguyen1,†, Tuyen P. Le1,†, Quan Nguyen3,‡

VinMotion1 VinUniversity2 USC3

*Equal contribution  ·  †Equal advising  ·  ‡Correspondence

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.

32%
Lower Normalized Peak Contact Force
78%
Lower Cumulative Peak Contact Force
70%
Fewer Unsafe-Contact Events
1.04±0.27 s
Mean Time-to-Stand
98%
Recovery Success Rate

Simulation, fallen-state protocol (N = 50). The first three are reductions vs. Passive Fall.

Abstract

Injury-aware falling and recovery

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.

Human ukemi demonstration (top) and humanoid ukemi-inspired falling and recovery (bottom).
Ukemi inspiration. A human performing a breakfall (top) and the humanoid performing an ukemi-inspired fall and recovery (bottom).
Condition-gated

Roll and recover

Rolling and recovery rewards switch on automatically from torso tilt and head height, so no separate phase estimator is needed.

Impact mitigation

Shape contact forces

Training penalizes heavy loading on the trunk and arms, which lowers peak contact force and the number of unsafe contacts.

Motion prior

Stay natural

A frozen Score-Matching Motion Prior scales the reward, keeping the learned motion close to natural human behavior.

Method

Condition-gated rewards with a motion prior

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.

Overview of Ukemi-SafeFall: inputs, frozen Score-Matching Motion Prior, condition-gated policy rewards, and evaluation metrics.
Overview of Ukemi-SafeFall. Inputs and disturbances feed a policy trained with gated task rewards and the frozen motion prior, then evaluated on injury-risk and recovery metrics.
Results

Comparison with baselines and ablations

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

Passive Fall

The controller outputs zero action, showing what an uncontrolled fall looks like.

Ukemi-SafeFall: Recovery Only

Trained only to get up, without rolling or impact-mitigation rewards.

Ukemi-SafeFall without Condition Gating

The ukemi rewards stay active for the whole episode instead of only when needed.

Ukemi-SafeFall without Score-Matching Motion Prior Reward Gating

Keeps the gated task rewards but removes the motion-prior reward gate.

SafeFall Triangle Proximal Policy Optimization (*)

A protective-fall baseline with no recovery objective. It uses a different training protocol, so it is a reference only.

Ukemi-SafeFall (Ours)

The full method: gated ukemi rewards, contact-force shaping and the motion-prior gate.

Sim-to-sim and hardware validation

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.

Simulation (sim-to-sim)

(a) Walking at 0.75 m/s over obstacles

(b) Running at 1 m/s over obstacles

(c) Descending stairs at 0.75 m/s

(d) Jumping from a 0.6 m-high platform

Hardware

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.

(a) Pushing from standing

(b) Kicking while walking

(c) Walking over a 0.08 m strip obstacle

(d) Jumping from a 0.6 m platform

Citation

Citation

@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.}
}