d3rlpy

Why we included this project

d3rlpy is a good fit when your reinforcement learning problem has to be solved from recorded data instead of live interaction. The library specializes in offline RL, learning a policy from a fixed dataset of past experiences, which matters in fields like robotics and healthcare where trying things out in the real world is costly or dangerous. It also keeps standard online training available through the same API, so you can switch between modes without learning a second tool. The implemented algorithms span classic choices like SAC and DQN, conservative methods like CQL, and newer transformer-based policies, all behind a config-driven interface that hides the PyTorch details. Tutorials and reproduction scripts help you check an implementation against published results before you trust it on your own problems.