Embodied AI Glossary中文Isaac Gym
CommonNVIDIA's GPU reinforcement-learning simulator launched in 2021; now discontinued.
Isaac Gym is a robot reinforcement-learning simulation platform NVIDIA released in 2021 (its technical report appeared in the NeurIPS 2021 Datasets and Benchmarks track), with physics computed by PhysX on the GPU. Its key feature is a tensor interface: physics results are handed to the neural network directly as PyTorch tensors without passing through the CPU, letting a single GPU run thousands of environments at once; NVIDIA claimed speedups of two to three orders of magnitude over a traditional “CPU simulation plus GPU training” pipeline. A wave of legged-locomotion and dexterous-hand work, including legged_gym, was built on it, and its companion repository IsaacGymEnvs shipped example environments such as Ant and ShadowHand. It was only ever released as a Preview build, with Preview 4 the final version; it is now discontinued, and NVIDIA recommends migrating to Isaac Lab, though many open-source projects still depend on it.
ExampleUnitree's unitree_rl_gym and ETH's legged_gym still require downloading and installing Isaac Gym Preview 3 or 4 before they can train walking policies for quadruped or humanoid robots.
- Also called
- NVIDIA Isaac Gym, IsaacGym, IsaacGymEnvs, Isaac Gym Preview 4
- Related
- NVIDIA Isaac Lab · PhysX · legged_gym · Massively Parallel Reinforcement Learning · NVIDIA Isaac Sim · Orbit / IsaacGymEnvs / OmniIsaacGymEnvs (predecessors of Isaac Lab)
- Sources
- Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning (arXiv 2108.10470)
NVIDIA Isaac Gym
isaac-sim/IsaacGymEnvs (GitHub) - As of
- 2026-09