SafeFall: Learning Protective Control for Humanoid Robots

TL;DR

SafeFall addresses the problem of protecting full-scale humanoid robots from fall-related damage by combining a lightweight GRU fall predictor with a damage-aware reinforcement learning policy that activates only when a fall is unavoidable. The method collects diverse failure trajectories, trains a predictor with conservative labeling, and learns a protection strategy via PPO using a multi-term reward that encodes component vulnerability and actuator limits. Key innovations include a two-stage curriculum, an asymmetric actor-critic framework, and domain randomization to enable sim-to-real transfer, demonstrated on a Unitree G1 where peak joint torque, contact force, and vulnerable-component collisions are dramatically reduced. Real-world experiments corroborate simulation results, showing high specificity, modest lead times, and substantial reductions in impulse and damage during omnidirectional falls. Overall, SafeFall provides a practical safety net that preserves nominal performance while enabling more aggressive experimentation and faster deployment of humanoid robots in unstructured environments.

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 \method, a framework that learns to predict imminent, unavoidable falls and execute protective maneuvers to minimize hardware damage. SafeFall is designed to operate seamlessly alongside 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 damage mitigation. The protective policy remains dormant until the predictor identifies a fall as unavoidable, at which point it activates to take control and execute a damage-minimizing response. This policy is trained with a novel, damage-aware reward function that incorporates the robot's specific structural vulnerabilities, learning to shield critical components like the head and hands while absorbing energy with more robust parts of its body. Validated on a full-scale Unitree G1 humanoid, SafeFall demonstrated significant performance improvements over unprotected falls. It reduced peak contact forces by 68.3\%, peak joint torques by 78.4\%, and eliminated 99.3\% of collisions with vulnerable components. By enabling humanoids to fail safely, SafeFall provides a crucial safety net that allows for more aggressive experiments and accelerates the deployment of these robots in complex, real-world environments.

Figures (6)

• Figure 1: SafeFall is the first method that protects full-scale humanoid robots from fall damages in the real world. The proposed SafeFall policy enables the humanoid to mitigate impact across a variety of scenarios, including (a) falling forward off the step, (b) falling backward, (c) rope-induced falls while running in 3 m/s, and (d) lateral falls. By prolonging the impact duration, distributing contact forces, and protecting fragile components, the policy achieves adaptive protection against complex, omnidirectional crashes.
• Figure 2: Overview of SafeFall. (a) We first collect diverse falling trajectories by rolling out the nominal locomotion policy under disturbances; (b) The trajectories are temporally segmented into safe, ambiguous, and falling phases to train a GRU-based fall predictor that identifies early signs of instability; (c) The SafeFall policy is trained using a series of damage‑mitigation rewards to safeguard joint motors and other fragile components. To enhance robustness, training proceeds from random initial states to progressively realistic and challenging falling states; (d) During deployment, the fall predictor operates alongside the nominal policy, adaptively switching control to the fall policy upon detecting an impending fall for effective protection.
• Figure 3: Illustration of critical components in a humanoid robot. The perception modules and dexterous manipulators are susceptible to physical impact. In contrast, the arms, legs, and torso shell act as protective structures, attenuating external impacts, preserving structural integrity, and shielding the embedded electronics. On the left, we visualize the difference between contact force, joint torque, and joint reaction force.
• Figure 4: Robot falling in different directions. Top: Lateral fall. The robot rotates its upper body to land on its high priority rigid torso, orienting its camera-equipped head upward while using its arms for auxiliary cushioning. Bottom: Forward fall. The robot extends its arms to brace for impact, absorbing energy and protecting its head from direct impact.
• Figure 5: Improvement in maximum contact force (left) and max joint force (right) achieved by SafeFall compared to the damping mode when the robot falls at various roll and pitch angles. SafeFall achieves significant improvement in most fall scenarios, highlighting its robust performance in omnidirectional falls.