frederikschubert/gwil

Re-Implementation of Cross-Domain Imitation Learning via Optimal Transport

The original paper is currently under review at ICLR 2022.

Core Idea

Minimising the Gromov-Wasserstein distance between the trajectory of an expert in one domain and an agent in a (possibly) different domain allows the agent to recover the expert's policy up to an isometry (rotation, translation, reflection).

Progress

• Added the GW calculation to a DQN agent on CartPole-v0 and try to immitate a static expert policy.
  • First experiments reach an average return of 50 after 7500 steps.
• Added the plain sinkhorn calculation and got better results immediately.