TorchRL is a Python-first reinforcement learning library built on PyTorch. It provides modular, composable primitives for building and training RL algorithms including support for multi-agent and model-based RL. The library targets researchers and practitioners working on decision-making systems in robotics and control environments.