acme

Why we included this project

Reinforcement learning researchers and practitioners often need to move from a quick laptop experiment to a distributed setup without rewriting their agent. Acme's core design separates acting from learning, so the same code runs in both modes. The library ships readable reference implementations of algorithms like DQN, R2D2, D4PG, and MPO, which serve as solid baselines and as starting points for new research. Because it builds on the DeepMind environment API and works with Reverb for replay, it fits naturally into JAX/TensorFlow workflows. The included examples and tutorials help bridge the gap between a paper and working code, which makes it a practical choice for students as well.