The paper discusses the challenges faced by autonomous robots when they encounter perception failures due to harsh environments or algorithm misinterpretations. The authors propose a model where such failures are treated as invisible obstacles, and a reinforcement learning-based local navigation policy is trained to guide the robot. The policy uses both proprioception and exteroception to sense collisions and react accordingly. Tested in real-time scenarios, this policy was found to increase the success rate by over 30% when facing perception failures, compared to existing heuristic-based locally reactive planners.
Publication date: 6 Oct 2023
Project Page: https://bit.ly/45NBTuh
Paper: https://arxiv.org/pdf/2310.03581
