GPU-Accelerated Parallel Simulation

Running thousands of simulated environments at once on a single GPU, shrinking robot training from days to minutes.

GPU-accelerated parallel simulation moves the physics simulation itself onto the GPU, advancing thousands of mutually independent environments at the same time and handing the simulation state directly to a framework like PyTorch as tensors for training — skipping the back-and-forth data copying between CPU and GPU that older pipelines needed. NVIDIA's Isaac Gym brought this approach into the mainstream in 2021; its paper claimed speedups of two to three orders of magnitude over the traditional “CPU simulation plus GPU training” pipeline. The most direct beneficiary is reinforcement learning: once sampling becomes cheap, policies for legged locomotion or dexterous manipulation can be trained on a single GPU in tens of minutes. Commonly used platforms today include Isaac Lab, MuJoCo MJX and MuJoCo Warp, Genesis, ManiSkill3, and Brax. The trade-off is that scenes must be written in a batched form, and complex contact or high-quality rendering noticeably slows things down.

ExampleETH Zürich's Rudin and colleagues simulated thousands of ANYmal quadruped robots in parallel on a single GPU, training a flat-ground walking policy in under 4 minutes and a rough-terrain policy in about 20 minutes, then transferred both successfully to the real robot.

Also called

Massively Parallel Simulation, GPU-Accelerated Simulation

Related

Vectorized Environments · Isaac Gym · NVIDIA Isaac Lab · Massively Parallel Reinforcement Learning · MuJoCo XLA · legged_gym

Sources

Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning (arXiv 2108.10470)
Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning (arXiv 2109.11978)