← Back ICRA 2026

Unified Humanoid Fall-Safety Policy from a Few Demonstrations

Zhengjie Xu, Ye Li, Kwan-Yee Lin, Stella Yu

PDF

AI summary

Key figure (auto-extracted from paper)

A single diffusion-driven policy trained on sparse human demonstrations enables humanoids to autonomously prevent falls, mitigate impact, and recover robustly across diverse disturbances.

Humanoid robotics Fall mitigation Reinforcement learning Diffusion policies Sim-to-real transfer Autonomous recovery

Problem

Prior humanoid control methods treat fall avoidance, impact mitigation, and recovery as isolated tasks, lacking an integrated strategy that handles the coupled dynamics of real-world falls and enables safe, autonomous recovery.

Approach

The authors fuse a few sparse human keyframe demonstrations with reinforcement learning and post-trajectory stitching to seed safe skills, then distill these into a diffusion-based adaptive memory that dynamically retrieves and composes safe fall-and-recovery trajectories in real time.

Key results

Robust sim-to-real transfer on a Unitree G1 humanoid
Significantly reduced impact forces during falls
Consistently fast and reliable autonomous recovery
Multi-modal safe reaction memory via diffusion policy

Why it matters

Enables safer, more resilient humanoid robots for real-world service and assistive applications by unifying fall safety into a single adaptive control framework.

Abstract

Falling is an inherent risk of humanoid mobility. Maintaining stability is thus a primary safety focus in robot control and learning, yet no existing approach fully averts loss of balance. When instability does occur, prior work addresses only isolated aspects of falling: avoiding falls, choreographing a controlled descent, or standing up afterward. Consequently, hu- manoid robots lack integrated strategies for impact mitigation and prompt recovery when real falls defy these scripts. We aim to go beyond keeping balance to make the entire fall-and-recovery process safe and autonomous: prevent falls when possible, reduce impact when unavoidable, and stand up when fallen. By fusing sparse human demonstrations with re- inforcement learning and an adaptive diffusion-based memory of safe reactions, we learn whole-body behaviors that unify fall prevention, impact mitigation, and rapid recovery in one policy. Experiments in simulation and on a Unitree G1 demonstrate robust sim-to-real transfer, lower impact forces, and consistently fast recovery across diverse disturbances, pointing toward safer, more resilient humanoids in real environments. Videos are available at https://firm2025.github.io/.

Index terms

Natural Machine Motion Whole-Body Motion Planning and Control Learning from Demonstration