In August 2021 a team led by Viktor Makoviychuk at NVIDIA published “Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning.” Isaac Gym is a platform for training robot control policies in which both the physics simulation and the neural-network training live on the GPU, eliminating the usual bottleneck of shuttling data back and forth to the CPU.
Traditional reinforcement-learning pipelines for robotics run the physics simulator on the CPU and the neural network on the GPU, so simulated experience constantly has to be copied between the two. Isaac Gym keeps everything on the GPU and passes data directly from physics buffers into PyTorch tensors. The authors reported speedups of two to three orders of magnitude compared with conventional CPU-simulator setups, making it possible to train complex robotics tasks on a single GPU.
This raw throughput is what made massively parallel locomotion training practical, and Isaac Gym became the substrate for a wave of legged-robot, manipulation, and dexterity results in the early 2020s. NVIDIA later folded its capabilities into the broader Isaac Sim and Isaac Lab tooling.
For a general reader, Isaac Gym is a good example of how a hardware and software optimization that sounds purely technical, keeping the whole training loop on one chip, can unlock entire research directions by making experiments cheap and fast.