This repository outlines the results of my master thesis on training a little robot in my own simulator and letting it interact with pedestrians controlled by Social Force.
The results show a very promising defensive policy that manages to avoid collisions entirely, given a very dense environment like in crowded pedestrian zones. This defensive policy can be combined with an offensive policy to maneuver around pedestrians on side-walks with only a few pedestrians.
- code/ contains the code of robot-sf and pysocialforce, as well as the (unfinished) dreamer implementation
- maps/ contains the training environments that were used to train the robot
- proofs/ contains a mathematical proof of the obstacle force as a virtual potential field
- results/ contains videos of the policies, trained agents, training logs and performance profiles
- demo/ contains the videos shown during the final presentation