Skip to main content
Anh Duc Tran

2026 / Robot learning

Ukemi-SafeFall

Teaching a humanoid robot to fall like a martial artist: rolling to soften the impact, then getting back up on its own.

Reinforcement learning · Humanoids · Safe falling · Motion priors · Sim-to-real

Co-first author

Paper project pagePaper · Humanoids 2026, accepted, to be presented

Real robot: jumping off a 0.6 m platform, rolling through the fall, and standing back up.

Key results

Compared with passive falling, in simulation (fallen-state protocol, 50 episodes per method)

lower normalized peak contact force
32%
lower cumulative peak contact force
78%
fewer unsafe-contact events
70%

Ukemi-SafeFall on its own, same protocol (absolute results)

recovery success rate
98%
mean time to stand
1.04± 0.27 s

These numbers are measured in simulation. The hardware clips show the behavior transferring to a real robot.

Goal, approach and outcome

01 / Goal
When a humanoid can't avoid a fall, reduce how hard it hits the ground and still let it stand back up afterward.
02 / Approach
Train a reinforcement learning policy with rewards inspired by martial-arts breakfalls, plus a learned motion prior that keeps the movements human-like.
03 / Outcome
The policy runs on a real 23-joint humanoid across pushes, kicks, trips, and drops. In simulation it cuts peak contact force by 32% and unsafe contacts by 70% compared with passive falling, and recovers from a fallen state in 98% of trials.

On the real robot

We deployed the policy on our in-house humanoid (23 actuated joints) and knocked it over in different ways. Each time it turns the fall into a roll, spreads the impact across its body, and stands back up on its own.

Pushed over from standing

Kicked while walking

Tripping over an 8 cm strip while walking

Longer take: platform and stairs

Method

Why falling matters

Most humanoid controllers treat a fall as the end of the episode. But real robots do fall. A push, a trip, or a misstep can knock them over, and a hard landing can damage the head, torso, or hips. That leaves an open question: if a fall is unavoidable, how should the robot fall?

Martial artists answer this with ukemi (breakfalls): instead of landing flat, they roll, spread the impact across the body, and come back up ready to move. We wanted the robot to learn the same habit.

Top: a human breakfall (ukemi). Bottom: the humanoid doing the same kind of roll and standing back up.
A person performing an aikido breakfall (top) and the humanoid performing a similar roll and recovery (bottom)

How it works

The robot learns by trial and error in simulation (reinforcement learning with PPO). Its rewards are split into three ideas:

  • Roll: start a roll and keep the momentum going instead of landing flat.
  • Soften the impact: limit how hard the torso and arms hit the ground.
  • Recover: get upright and raise the head back to standing height.

These rewards switch on automatically based on how tilted the body is and how low the head is, so the robot doesn't need a separate module to decide which phase of the fall it's in.

On top of that, a diffusion-based Score-Matching Motion Prior, trained on human motion, scores how natural each movement looks. The task reward is multiplied by that score, so the robot only gets full credit for motions that are both safe and human-like. The same prior is also used to generate varied fallen starting poses for training.

Framework overview
Ukemi-SafeFall framework: inputs, frozen motion prior, condition-gated rewards, and evaluation

Results

The headline numbers are above. These clips show where they come from: a controlled comparison, an ablation of each part of the method, and further simulated falls.

Controlled comparison: passive fall vs. our policy

Same push, same starting pose, in simulation. Without control the robot just collapses. With Ukemi-SafeFall it rolls through the fall and stands back up.

Passive fall

No control. This is the baseline for uncontrolled falling.

Ukemi-SafeFall (ours)

Rolls to spread the impact, then recovers to standing.

One fall, frame by frame. Red lines show contact forces.
Simulation frames of the humanoid rolling through a fall and standing up, with contact forces shown in red

Ablations: what each piece adds

Removing parts of the method shows why each one is there.

Recovery only

Trained only to stand up, with no rolling or impact rewards.

Without condition gating

All rewards stay on for the whole episode instead of switching on at the right moment.

Without the motion prior

Safe-falling rewards without the human-likeness score.

Additional simulation experiments

Before going to hardware, we tested the policy in a separate MuJoCo simulator on the same kinds of falls.

Walking at 0.75 m/s over 8 cm obstacles

Running at 1 m/s over 8 cm obstacles

Falling while descending stairs

Jumping from a 0.6 m platform

Paper

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

Anh Duc Tran*, Anh-Quang Vu*, Loc Pham, Trong Hieu Nguyen, Chuong Nguyen Le, Quang Huy Dao, Hoang Vu Dao, Van-Truong Nguyen, Chuong Nguyen, Pham Tuyen Le, Quan Nguyen

* Equal contribution

IEEE-RAS International Conference on Humanoid Robots (Humanoids), 2026. Accepted, to be presented.

Paper project page

Cite

@inproceedings{tran2026ukemisafefall,
  title     = {Ukemi-SafeFall: An Ukemi-inspired Injury-Aware Falling and Recovery},
  author    = {Anh Duc Tran and Anh-Quang Vu and Loc Pham and Trong Hieu Nguyen and Chuong Nguyen Le and Quang Huy Dao and Hoang Vu Dao and Van-Truong Nguyen and Chuong Nguyen and Pham Tuyen Le and Quan Nguyen},
  booktitle = {IEEE-RAS International Conference on Humanoid Robots (Humanoids)},
  year      = {2026},
  note      = {Accepted, to be presented. Anh Duc Tran and Anh-Quang Vu contributed equally.}
}

Next project

2025 / Robot learning

Reinforcement Learning for Quadruped Walking

Training a four-legged robot to walk smoothly across different terrain in simulation, then deploying the learned controller on a real Unitree Go1.