Teaching a Machine to Read Maps with Deep Reinforcement Learning Gino Brunner, Oliver Richter, Yuyi Wang, Roger Wattenhofer ETH Zurich
Our map Our view Deepmind randomly generated, starting on small maps
Reinforcement Learning Environment Reward State Action Finding the target Bumping into walls Compass Agent
What does it mean to read a map? Localize
What does it mean to read a map? Localize Find a path to the target Follow the path I think I’m here
t + 1 Visual Input I think I’m here Estimated Position Map Input Compass Visible Local Map Network Recurrent Localization Cell Estimated Position I think I’m here Map Input Policy π Location Uncertainty Acting Agent Map Interpretation Network Short Term Target Direction
Modular design as key to success Visible Local Map Network Map Interpretation Network Acting Agent Recurrent Localization Cell Estimated Position Reward Prediction Policy π Actual Position Actual Reward Reinforcement Learning Neural nets, heuristics, explicit algorithms
Results
Questions? Oliver Richter richtero@ethz.ch
We and robots navigate with maps and gps Or SLAM >no sensors but map> human
Visible Local Map Network FC Orientation Visual Input Visible Local Map FC Visible Field Map Excerpt CNN FC