Research Analyzer
← Back ICRA 2026

Robot Crash Course: Learning Soft and Stylized Falling

Pascal Strauch, David Müller, Sammy Christen, Agon Serifi, Ruben Grandia, Espen Knoop, Moritz Bächer

PDF

AI summary

Key figure (auto-extracted from paper)
A reinforcement learning policy enables bipedal robots to execute controlled, soft falls while allowing users to specify arbitrary stylized end poses for artistic or recovery purposes.
Reinforcement learning soft falling legged robots stylized motion sim-to-real impact mitigation

Problem

Traditional fall mitigation strategies either restrict robot capabilities to prevent falls or result in uncontrolled, damaging impacts without user control over the final pose.

Approach

The authors train a reinforcement learning policy with a custom reward function that balances impact minimization and user-specified end-pose tracking, supported by a physics-informed sampling strategy for diverse initial and final states.

Key results

  • Reduces peak and mean impact forces compared to standard falling strategies
  • Enables accurate tracking of unseen, user-specified stylized end poses at inference
  • Demonstrates successful sim-to-real transfer with zero damage on a custom bipedal robot
  • Provides a tunable trade-off between damage reduction and pose accuracy via adjustable reward weights

Why it matters

It transforms robot falls from catastrophic failures into controllable, lifelike motions, benefiting both human-robot interaction and robust recovery strategies for legged systems.

Abstract

Despite recent advances in robust locomotion, bipedal robots operating in the real world remain at risk of falling. While most research focuses on preventing such events, we instead concentrate on the phenomenon of falling itself. Specifically, we aim to reduce physical damage to the robot while providing users with control over the robot’s end pose. To this end, we propose a robot-agnostic reward function that balances the achievement of a desired end pose with impact minimization and the protection of critical robot parts during reinforcement learning. To make the policy robust to a broad range of initial falling conditions, and to enable the specification of an arbitrary and unseen end pose at inference time, we introduce a simulation-based sampling strategy of initial and end poses. Through simulated and real-world experiments, our work demonstrates that even bipedal robots can perform controlled, soft falls.

Index terms

Humanoid and Bipedal Locomotion Reinforcement Learning Machine Learning for Robot Control

Related papers