Legged Gym: A Framework for Massively Parallel Reinforcement Learning of Legged Locomotion

Legged Gym emerged from a simple but powerful observation: deep reinforcement learning for legged locomotion does not need to take days on server clusters if the simulation itself is massively parallelized on a single GPU. Built as a high-level wrapper around NVIDIA Isaac Gym, Legged Gym provides a complete training stack—including environment definitions, reward functions, curriculum strategies, and policy architectures—for training quadruped and humanoid robots in simulation. By running thousands of environments simultaneously on a single GPU, the framework compresses training times from days to minutes without sacrificing policy quality. The framework ships with environments for several popular robot platforms, including Unitree A1, Go1, and ANYmal, as well as the bipedal robot Cassie. Each environment includes a carefully tuned reward function and terrain curriculum that progressively increases task difficulty as the policy improves. The terrain curriculum is particularly noteworthy: it is inspired by game design principles, procedurally generating increasingly challenging elevation maps, stairs, and obstacles to prevent premature convergence and encourage robust gaits. The authors show that policies trained with this curriculum transfer directly to real-world hardware with minimal domain randomization, a testament to the physical fidelity of the Isaac Gym simulator and the robustness of the learned gaits. From a software engineering perspective, Legged Gym is clean, modular, and highly extensible. Environment parameters are managed through structured YAML configs, making it easy to sweep reward weights, terrain properties, and robot dynamics. The training loop uses PyTorch and leverages Isaac Gym’s GPU-based physics backend, so all simulation rollouts happen on the GPU with minimal CPU-GPU transfer overhead. The framework also supports real-time visualization, logging, and checkpointing, enabling rapid iteration and debugging. Perhaps the most impactful result is the training speed benchmark. On a single NVIDIA RTX 3090, Legged Gym trains a flat-terrain walking policy for ANYmal in under four minutes, and an uneven-terrain policy in under twenty minutes. This represents orders of magnitude speedup compared to traditional CPU-based simulation pipelines, democratizing access to legged robot learning for labs without large compute budgets. The codebase has become a de facto standard in the legged locomotion community, with hundreds of derivative works extending it to new robots, new tasks, and new learning algorithms.