rl

Why we included this project

TorchRL approaches reinforcement learning the PyTorch way. Instead of one fixed framework, you get composable pieces: environments, transforms, collectors, replay buffers, and loss functions that you wire together into your own training pipeline. The code stays close to how you already write PyTorch, which helps when you are testing a new algorithm or shipping a control agent and want to swap components without fighting an abstraction layer. Wrappers for Gym, DMControl, Brax, and others, plus native multi-agent support, cover both single-agent and distributed setups. Being part of the PyTorch ecosystem also means solid documentation, examples, and steady maintenance, so a small team can adopt it and keep building on it.