SafeFall:
Learning Protective Control for Humanoid Robots

Ziyu Meng1,2,*, Tengyu Liu2,*, Le Ma2, Yingying Wu2,3, Ran Song1,†, Wei Zhang1, Siyuan Huang2,†

(*: equal contribution, †: corresponding author)

1School of Control Science and Engineering, Shandong University 2National Key Laboratory of General Artificial Intelligence, BIGAI 3Department of Automation, Tsinghua University
Abstract.

Bipedal locomotion makes humanoid robots inherently prone to falls, causing catastrophic damage to the expensive sensors, actuators, and structural components of full-scale robots. To address this critical barrier to real-world deployment, we present SafeFall, a framework that learns to predict imminent, unavoidable falls and execute protective maneuvers to minimize hardware damage. SafeFall is designed to operate seamlessly alongside any existing nominal controller, ensuring no interference during normal operation. It combines two synergistic components: a lightweight, GRU-based fall predictor that continuously monitors the robot's state, and a reinforcement learning policy for mitigation.

Approach (Overview)
SafeFall Approach Overview

SafeFall combines a lightweight GRU-based fall predictor with a damage-aware reinforcement learning policy that executes protective maneuvers when a fall is predicted to be unavoidable. The predictor continuously monitors the robot state and, upon detection, the mitigation policy activates to minimize hardware damage by shielding vulnerable components while distributing impact to more robust ones.

Video.
Data collection.

We simulate various fall-inducing scenarios in the simulation to collect data, which is subsequently used to train the fall predictor and initialize the starting poses for the SafeFall policy.

Real-world Demo.
Nominal Scenarios

The fall predictor yields no false positives for disturbances that fall within the controllable range of the nominal policy.

Fall from different directions

The SafeFall policy is capable of handling omnidirectional falls.

More challenging scenarios

The SafeFall policy maintains its protective capabilities even in more challenging scenarios.

Citation

BibTeX
@article{meng2026safefall,
        title     = {SafeFall: Learning Protective Control for Humanoid Robots},
        author    = {Ziyu Meng and Tengyu Liu and Le Ma and Yingying Wu and Ran Song and Wei Zhang and Siyuan Huang},
        journal   = {arXiv preprint},
        year      = {2026}
}