Event

OSCAR: A Navigation Framework that Learns Waiting Times by Obstacle

1 reports · 1 sources

Hshmat Sahak et al. propose the OSCAR framework, which enables a robot to learn remaining clearance-time distributions from online experience according to obstacle category, integrate them into a time-varying graph planner with obstacle memory, and decide on each blocked edge how long to wait before detouring. The method assumes that the obstacle's category label is available when the obstacle is encountered. The paper claims that in simulation, after fewer than 20 observations per obstacle category, the learned policy's time to reach the goal converges to within 1% of an oracle that knows the true clearance distribution, outperforming all heuristic baselines; on a real mobile robot in a university atrium with obstacles including people, chairs, trash cans, and pipes, the policy adjusts its patience threshold online across 50 navigation episodes.