fast-nav · point-nav policy on held-out scenes

noise
speed
goals 0/0 steps drift est. dist inference
size
policy input
full observation: 64 lidar ranges (6 m max), odometry frame
goal vector · previous action
loading…
click: set goal · shift-click: teleport · drag: pan · wheel / pinch: zoom · wall / erase: edit the map
hollow ring = robot's odometry estimate (visible with noise on)
GRU-256 PPO policy, ~98% success on held-out scenes · sim + inference in JS
loading policy + scenes…