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

Overview

Injury-aware falling and recovery

Humanoid robots remain vulnerable to falls when balance recovery fails. Controllers that treat falling as terminal failure leave an open safety problem: how to reduce mechanical damage during unavoidable falls while preserving the option to stand up afterward.

This paper proposes Ukemi-SafeFall, an ukemi-inspired, injury-aware falling and recovery framework that shapes humanoid fall–recovery behavior through condition-gated rolling and impact-mitigation rewards. In simulation, Ukemi-SafeFall reduces injury-risk surrogates of normalized peak contact force, cumulative peak contact force, and unsafe-contact events by 32%, 78%, and 70%, respectively, compared with passive falling. Experiments further demonstrate that Ukemi-SafeFall transfers to a humanoid platform across a range of falling scenarios.

On a fallen-state evaluation protocol (50 simulated episodes per method), Ukemi-SafeFall reaches a 98% recovery success rate with time-to-stand 1.04±0.27 s. The same protocol is used for matched ablations and baselines described below.

Human ukemi demonstration (top) and humanoid ukemi-inspired falling and recovery (bottom). — **Ukemi inspiration.** *Ukemi* denotes martial-arts breakfall techniques intended to reduce hazardous loading during a fall. A humanoid robot (bottom) performs ukemi-inspired falling and recovery maneuvers for unavoidable falls, inspired by ukemi techniques demonstrated by a human (top).

Condition-gated ukemi

Condition-gated roll and recovery

Rewards activate from torso tilt and head height to encourage roll initiation, roll continuation, soft landing, and standing recovery, without a discrete phase estimator.

Impact mitigation

Contact-force shaping

Training limits trunk and arm loading during contact so normalized peak contact force, cumulative peak contact force, and unsafe-contact events are reduced relative to uncontrolled falling.

Recovery-oriented

Standing recovery with motion-prior gating

Recovery rewards encourage upright posture and head-height progress. A frozen Score-Matching Motion Prior multiplies the task reward to keep motions consistent with human reference behavior.

What we measure

We report injury-risk surrogates (simulation contact quantities, not direct hardware damage) and recovery surrogates under the fallen-state protocol. Names match the paper definitions.

Normalized peak contact force — largest contact force in an episode, expressed as a multiple of body weight. Lower indicates less severe peak impact.
Cumulative peak contact force — accumulated peak contact loading over time during non-foot contact, reflecting both magnitude and duration.
Unsafe-contact events — count of potentially hazardous head-proxy, torso, and pelvis contacts.
Time-to-stand — time from the first non-foot ground contact to first satisfaction of the standing criteria (reported as mean ± std. over successful episodes).
Recovery success rate — fraction of episodes that meet the standing criteria at the final evaluation step.

Framework

Method overview

We train a reinforcement learning policy with Proximal Policy Optimization for humanoid falling and recovery. The policy maps partial observations to target joint offsets. Training uses a multiplicative reward that combines condition-gated task rewards with a diffusion-based Score-Matching Motion Prior reward. Generative State Initialization samples diverse fallen starting poses from the same prior, and domain randomization varies physical and sensing parameters so the policy sees a wide range of conditions.

Overview of Ukemi-SafeFall: inputs, frozen Score-Matching Motion Prior, condition-gated policy rewards, and evaluation metrics. — **Overview of Ukemi-SafeFall.** A reinforcement learning framework trained via Proximal Policy Optimization for humanoid falling and recovery. The policy uses a multiplicative reward that combines condition-gated task rewards (safe rolling and impact mitigation) with a diffusion-based Score-Matching Motion Prior reward. Generative State Initialization samples fallen starts from the same prior.

Comparisons

Comparisons in simulation

Using the same fallen-state evaluation protocol, we evaluate Ukemi-SafeFall under three ablations: Ukemi-SafeFall: Recovery Only, Ukemi-SafeFall without condition gating, and Ukemi-SafeFall without Score-Matching Motion Prior reward gating. We also compare against Passive Fall and, as a cross-protocol reference, SafeFall Triangle Proximal Policy Optimization.

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

Method rollouts

Illustrative simulation rollouts for each comparison method and ablation.

Passive Fall

The controller outputs zero action after Generative State Initialization and is subjected to the evaluation disturbances. This baseline characterizes uncontrolled falling.

Ukemi-SafeFall: Recovery Only

This ablation uses the same multiplicative Score-Matching Motion Prior reward gate and Generative State Initialization procedure as Ukemi-SafeFall, but it is trained only with the upward-motion and head-height recovery rewards.

Ukemi-SafeFall without Condition Gating

This ablation retains the proposed ukemi-inspired rewards, contact-force shaping, and Score-Matching Motion Prior gate, but disables the activation conditions; the corresponding reward terms remain active throughout the episode.

Ukemi-SafeFall without Score-Matching Motion Prior Reward Gating

This ablation retains the condition-gated task rewards and contact-force shaping, but removes the multiplicative Score-Matching Motion Prior reward gate. Generative State Initialization continues to sample initial states from the Score-Matching Motion Prior.

SafeFall Triangle Proximal Policy Optimization (*)

We re-implement a SafeFall method to show how a protective falling policy performs without an explicit recovery objective. It is adapted to the same humanoid but uses a different training protocol; therefore, we report it as a cross-protocol reference rather than a matched ablation.

Ukemi-SafeFall (Ours)

The complete method combines condition-gated ukemi-inspired rewards, contact-force shaping, and multiplicative Score-Matching Motion Prior reward gating.

Results

Sim-to-sim and hardware validation

We validate Ukemi-SafeFall across falling scenarios including locomotion over obstacles, stair descent, and platform jumps. In a separate MuJoCo-based sim-to-sim environment, the controller exhibits rolling, distributed limb contact, and recovery toward standing. Hardware demonstrations are coming soon.

Simulation

(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

Sim-to-sim results: (a) walking and (b) running over 0.08 m-high obstacles, (c) descending stairs, and (d) jumping from a 0.6 m-high platform.

Hardware

Hardware experiments Coming soon