Unified Humanoid Fall-Safety Policy from a Few Demonstrations

TL;DR

FIRM tackles the inherent risk of humanoid falls by unifying fall damage mitigation and recovery into a single policy learned from sparse human demonstrations. It combines seed skill priors, augmentation via reinforcement learning and post-trajectory stitching, and a diffusion-based adaptive memory with an online adapter to handle diverse disturbances. In simulation and on a Unitree G1, FIRM achieves lower impact forces, robust sim-to-real transfer, and faster recoveries across multiple terrains and payloads, outperforming state-of-the-art baselines. This work advances safe, resilient humanoid control for unstructured environments, while noting limitations related to memory-driven nearest-neighbor retrieval and proprioception-only sensing.

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, humanoid 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 reinforcement learning and an adaptive diffusion-based memory of safe reactions, we learn adaptive 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 towards safer, more resilient humanoids in real environments. Videos are available at https://firm2025.github.io/.

Figures (12)

• Figure 1: Our method enables humanoids to fall safely and rise promptly. Snapshots show real-world deployment on the Unitree G1: When suddenly destabilized, the robot redirects into a side fall with arm buffering, then reorients and rises, demonstrating adaptive and resilient recovery.
• Figure 2: Workflow Overview. From a few sparse human key poses, the robot seeds safe fall–rise skills, expands them through RL with post-trajectory stitching, distills enriched behaviors into a diffusion-based action memory, and composes online adapter to execute actions with context-awareness.
• Figure 3: Overview of online adapter. During inference, the adapter uses the history of observations to dynamically predict a feature and match with a key-frame goal feature in the code-book, and then pass the matched goal feature into the diffusion model to guide the process with context-awareness.
• Figure 4: (a) Distribution of contact force on the base over all time steps. Time steps with base contact impulse below $0.05\,\mathrm{Ns}$ are not included. (b) Base acceleration (BA) during the fall.
• Figure 5: Motion behaviors under different payloads. As the white boxes show, For 2 kg payload, the arms perform a “support–push” motion. For 12 kg, the robot’s arms make full contact with the ground, exhibiting a forceful pushing action to lift the body. The orange arrow indicates torso orientation.