LaND: Learning to Navigate From Disengagements

Consistently testing autonomous mobile robots in real world scenarios is a necessary aspect of developing navigation systems. Each time the human safety monitor disengages robot's autonomy system due to robot performing an undesirable maneuver, developers gain insight into how improve system. However, we believe that these disengagements not only show where fails, which useful for troubleshooting, but also provide direct learning signal by can learn navigate. We present reinforcement approach navigate from disengagements, or LaND. LaND learns neural network model predicts actions lead given current sensory observation, and then at test plans executes avoid disengagements. Our results demonstrate successfully diverse, sidewalk environments, outperforming both imitation approaches. Videos, code, other material are available on our website https://sites.google.com/view/sidewalk-learning.