protomotions.agents.ase package

Adversarial Skill Embeddings (ASE) implementation.

This package implements the ASE algorithm which extends AMP with learned skill embeddings. The discriminator encodes motions into a latent space and the policy is conditioned on these latent codes for diverse skill learning.

Key Components:

  • ASE: Main ASE agent

  • ASEModel: Actor-critic with skill encoder

  • ASEAgentConfig: Configuration

Example

>>> from protomotions.agents.ase.agent import ASE
>>> agent = ASE(fabric, env, config)
>>> agent.train()
Reference:

Peng et al. “ASE: Large-Scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters” (2022)

Submodules