This repo forks the original NavRep repo mainly developed by Daniel Dugas: https://github.com/ethz-asl/navrep
This work explores three avenues aiming to located ways of improving navigation success rate in realistic crowded environments. The three avenues are:
- Pretrained initialisation based on behaviour cloning of an expert policy,
pretrain - Intrinsic curiosity driven learning,
reward - Curriculum based learning using more realistic maps,
curriculum
The approaches are explored independently on different branches, linked above. The individual READMEs explain how to use the scripts.
This library was written primarily by Daniel Dugas. The transformer block codes, and vae/lstm code were taken or heavily derived from world models and karpathy's mingpt. We've retained the copyright headers for the relevant files.