REBot: Reflexive Evasion Robot for Instantaneous Dynamic Obstacle Avoidance
ZIHAO XU, Ce Hao, Chunzheng Wang, Kuankuan Sima, Fan Shi, Jin Song Dong
AI summary
Problem
Existing dynamic obstacle avoidance methods rely on trajectory replanning that requires over two seconds of reaction time, causing quadrupedal robots to fail when obstacles approach rapidly. Legged platforms currently lack general-purpose frameworks for instantaneous, reflexive evasion under constrained reaction times.
Approach
REBot employs a finite-state machine that transitions between normal, avoidance, and recovery stages based on obstacle proximity and robot stability. It trains reinforcement learning policies using a two-stage curriculum, regularization, and adaptive rewards to execute rapid, balance-preserving evasive maneuvers without slow navigation replanning.
Key results
- Highest avoidance and recovery success rates across static and dynamic obstacle scenarios in simulation and real-world tests
- Reduced maximum joint power and avoidance distance compared to baseline navigation-based and reactive RL methods
- Successful real-time evasion on a physical Unitree Go2 robot against fast-moving threats with reaction times under 1.5 seconds
- Ablation studies confirm curriculum learning, adaptive rewards, and the dedicated recovery policy significantly boost evasion reliability
Why it matters
Provides a critical safety framework for deploying legged robots in fast-changing, human-populated environments where split-second reflexive reactions are necessary.
Abstract
Dynamic obstacle avoidance (DOA) is critical for quadrupedal robots operating in environments with moving obstacles or humans. Existing approaches typically rely on navigation-based trajectory replanning, which assumes sufficient reaction time and leading to fails when obstacles approach rapidly. In such scenarios, quadrupedal robots require reflexive evasion capabilities to perform instantaneous, low-latency ma- neuvers. This paper introduces Reflexive Evasion Robot (REBot), a control framework that enables quadrupedal robots to achieve real-time reflexive obstacle avoidance. REBot integrates an avoid- ance policy and a recovery policy within a finite-state machine. With carefully designed learning curricula and by incorporating regularization and adaptive rewards, REBot achieves robust evasion and rapid stabilization in instantaneous DOA tasks. We validate REBot through extensive simulations and real-world experiments, demonstrating notable improvements in avoidance success rates, energy efficiency, and robustness to fast-moving obstacles. Paper homepage: https://rebot-2025.github.io/.