Reproduce “REBot: Reflexive Evasion Robot for Instantaneous Dynamic Obstacle Avoidance” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zihao Xu, Ce Hao, Chunzheng Wang, Kuankuan Sima, Fan Shi, Jin Song Dong
A control framework called REBot enables quadrupedal robots to perform instantaneous, reflexive evasive maneuvers against dynamic obstacles with severely limited reaction times. It achieves avoidance success rates of 65% in simulation for 0.5-1.5s reaction times and demonstrates real-world evasion against various threats, outperforming traditional navigation-based methods.
Key insights
- Reflexive evasion, inspired by biological spinal reflexes, is a necessary and effective strategy for quadrupedal robots to avoid fast-approaching obstacles when deliberation and replanning are not feasible.
- Incorporating adaptive reward terms during reinforcement learning training, such as those encouraging gait diversity and threat-level response, is crucial for developing robust and versatile evasive behaviors.
- A specific recovery policy is essential to ensure post-evasion stability, particularly for aggressive, low-latency maneuvers that can compromise the robot's balance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.