Isaac Gym is a high-performance, GPU-accelerated robotics simulation platform designed to train policies for various robotics tasks. It integrates physics simulation and neural network policy training directly on GPUs, eliminating the need for data transfer between CPU and GPU, which significantly reduces training times. This approach achieves 2-3 orders of magnitude improvements over conventional RL training methods that use CPU-based simulators and GPUs for neural networks. The platform supports a wide range of complex robotic environments, including locomotion, character animation, and manipulation tasks. Key contributions include a high-fidelity GPU-accelerated simulator, a Tensor API for direct access to physics buffers, and the ability to simulate hundreds of thousands of steps per second on a single GPU. Empirical results demonstrate significant speed-ups in training various simulated environments, with notable improvements in tasks such as ant locomotion, humanoid walking, and robotic hand manipulation. The platform also supports sim-to-real transfer, showcasing its ability to generalize learned policies to real-world scenarios.Isaac Gym is a high-performance, GPU-accelerated robotics simulation platform designed to train policies for various robotics tasks. It integrates physics simulation and neural network policy training directly on GPUs, eliminating the need for data transfer between CPU and GPU, which significantly reduces training times. This approach achieves 2-3 orders of magnitude improvements over conventional RL training methods that use CPU-based simulators and GPUs for neural networks. The platform supports a wide range of complex robotic environments, including locomotion, character animation, and manipulation tasks. Key contributions include a high-fidelity GPU-accelerated simulator, a Tensor API for direct access to physics buffers, and the ability to simulate hundreds of thousands of steps per second on a single GPU. Empirical results demonstrate significant speed-ups in training various simulated environments, with notable improvements in tasks such as ant locomotion, humanoid walking, and robotic hand manipulation. The platform also supports sim-to-real transfer, showcasing its ability to generalize learned policies to real-world scenarios.