Avalon-Benchmark/avalon

Procedurally generates 20 diverse tasks with identical action spaces and reward functions, enabling systematic evaluation of agent generalization across skills like navigation, hunting, and foraging. Built on a custom Godot engine optimized for headless GPU rendering on Linux, with OpenAI Gym integration and baseline implementations (PPO, Dreamer) included. Benchmarked against hundreds of hours of human gameplay data, revealing significant performance gaps that challenge existing RL algorithms.

190 stars. No commits in the last 6 months.