Reproduce “Robust Imitation of Diverse Behaviors” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ziyu Wang, Josh Merel, Scott Reed, Greg Wayne, Nando de Freitas, Nicolas Heess

The method combines a variational autoencoder (VAE) with Generative Adversarial Imitation Learning (GAIL) to condition a controller on demonstrated behavior. A trajectory encoder provides a behavior representation, while adversarial fine-tuning trains the policy through interaction with simulation. In simulated locomotion, fine-tuning reduced humanoid falls and favored smaller walker speed errors measured on training trajectories; some test styles overlapped with training.

Key insights

  • Conditioning the GAIL discriminator on VAE-learned semantic embeddings (z) effectively alleviates mode collapse, enabling the policy to capture the full diversity of demonstrated behaviors.
  • The VAE component successfully learns a structured, continuous latent space, allowing for smooth interpolation between different behaviors and supporting one-shot imitation of novel trajectories.
  • The combined VAE-GAIL framework synergistically leverages VAE's ability to model diverse distributions and GAIL's capacity for learning robust policies, overcoming individual method limitations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.