Toni-SM/skrl

Provides composable agents and algorithms (policy gradient, Q-learning, actor-critic methods) with transparent, readable implementations across multiple backends. Enables multi-scope training where distinct agent groups can simultaneously learn within isolated or shared environment subsets—critical for complex multi-agent and curriculum learning scenarios. Includes pre-trained model hosting on Hugging Face and unified APIs for diverse environment types from simulation platforms to robotics frameworks.

1,011 stars. Actively maintained with 4 commits in the last 30 days. Available on PyPI.