Action aided Joint Embedding Prediction on 2D LiDAR Sequences for Robust Mobile Robot Navigation in Dynamic Environments

Mobile robots that navigate among moving obstacles must act on 2D LiDARscans that are sparse, noisy, and change from one step to the next.Reinforcement learning policies such as PPO are usually trained on theseraw scans directly. A single scan, however, does not show how obstaclesare moving, nor how much of the change between consecutive scans comesfrom the robot's own motion. The policy has to learn both perception andcontrol from a sparse reward, and in our baseline experiments this madeperformance sensitive to sensor noise and highly variable across randomseeds. Supervising a better representation would require labels forobstacle motion, which are costly to collect outside simulation. Wetherefore ask whether a self-supervised model, trained only on unlabelledLiDAR sequences and the robot's actions, can learn a representation ofobstacle dynamics that improves a downstream PPO policy's success rateunder sensor noise and reduces its variability across seeds.