Isaac Gym: High Performance GPU-Based Physics Simulation For Robot\n Learning

Abstract

Isaac Gym offers a high performance learning platform to train policies for\nwide variety of robotics tasks directly on GPU. Both physics simulation and the\nneural network policy training reside on GPU and communicate by directly\npassing data from physics buffers to PyTorch tensors without ever going through\nany CPU bottlenecks. This leads to blazing fast training times for complex\nrobotics tasks on a single GPU with 2-3 orders of magnitude improvements\ncompared to conventional RL training that uses a CPU based simulator and GPU\nfor neural networks. We host the results and videos at\n\\url{https://sites.google.com/view/isaacgym-nvidia} and isaac gym can be\ndownloaded at \\url{https://developer.nvidia.com/isaac-gym}.\n