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 LiDAR
scans that are sparse, noisy, and change from one step to the next.
Reinforcement learning policies such as PPO are usually trained on these
raw scans directly. A single scan, however, does not show how obstacles
are moving, nor how much of the change between consecutive scans comes
from the robot's own motion. The policy has to learn both perception and
control from a sparse reward, and in our baseline experiments this made
performance sensitive to sensor noise and highly variable across random
seeds. Supervising a better representation would require labels for
obstacle motion, which are costly to collect outside simulation. We
therefore ask whether a self-supervised model, trained only on unlabelled
LiDAR sequences and the robot's actions, can learn a representation of
obstacle dynamics that improves a downstream PPO policy's success rate
under sensor noise and reduces its variability across seeds.