Isaac Gym is a high-performance GPU-based physics simulation platform for robot learning. It enables training of policies for a wide variety of robotics tasks directly on GPU, with both physics simulation and neural network policy training running on GPU and communicating by directly passing data from physics buffers to PyTorch tensors without any CPU bottlenecks. This leads to significantly faster training times for complex robotics tasks on a single GPU, with improvements of 2-3 orders of magnitude compared to conventional RL training that uses a CPU-based simulator and GPU for neural networks. Isaac Gym provides a PyTorch tensor-based API to access the results of physics simulation natively on the GPU, allowing observation tensors to be used as inputs to a policy network and action tensors to be directly fed back into the physics system. It supports a wide range of robotic environments, including locomotion environments, robotic hands, and humanoid character animation. The platform provides a simple procedural API to create environments and populate them with actors, supporting loading data from URDF and MJCF file formats. It also includes a basic Proximal Policy Optimization (PPO) implementation and a straightforward RL task system, but users may substitute alternative task systems or RL algorithms as desired. Isaac Gym leverages NVIDIA PhysX to provide a GPU-accelerated simulation back-end, allowing it to gather experience data required for robotics RL at rates only achievable using a high degree of parallelism. The platform provides a Tensor API that allows users to interact with the running simulation on either CPU or GPU, enabling efficient training of policies for complex robotics tasks. The results show that Isaac Gym achieves significant speed-ups in training various simulated environments, with Ant and Humanoid environments achieving performant locomotion in 20 seconds and 4 minutes respectively, ANYmal in under 2 minutes, Humanoid character animation using AMP in 6 minutes, and cube rotation with Shadow Hand in 35 minutes all on a single NVIDIA A100 GPU. The platform also demonstrates sim-to-real transfer results on ANYmal and TriFinger, showcasing the ability of the simulator to perform high-fidelity contact-rich manipulation. The results show that Isaac Gym provides a high-performance training platform for robotics, with significant improvements in training speed and efficiency compared to conventional RL training methods.Isaac Gym is a high-performance GPU-based physics simulation platform for robot learning. It enables training of policies for a wide variety of robotics tasks directly on GPU, with both physics simulation and neural network policy training running on GPU and communicating by directly passing data from physics buffers to PyTorch tensors without any CPU bottlenecks. This leads to significantly faster training times for complex robotics tasks on a single GPU, with improvements of 2-3 orders of magnitude compared to conventional RL training that uses a CPU-based simulator and GPU for neural networks. Isaac Gym provides a PyTorch tensor-based API to access the results of physics simulation natively on the GPU, allowing observation tensors to be used as inputs to a policy network and action tensors to be directly fed back into the physics system. It supports a wide range of robotic environments, including locomotion environments, robotic hands, and humanoid character animation. The platform provides a simple procedural API to create environments and populate them with actors, supporting loading data from URDF and MJCF file formats. It also includes a basic Proximal Policy Optimization (PPO) implementation and a straightforward RL task system, but users may substitute alternative task systems or RL algorithms as desired. Isaac Gym leverages NVIDIA PhysX to provide a GPU-accelerated simulation back-end, allowing it to gather experience data required for robotics RL at rates only achievable using a high degree of parallelism. The platform provides a Tensor API that allows users to interact with the running simulation on either CPU or GPU, enabling efficient training of policies for complex robotics tasks. The results show that Isaac Gym achieves significant speed-ups in training various simulated environments, with Ant and Humanoid environments achieving performant locomotion in 20 seconds and 4 minutes respectively, ANYmal in under 2 minutes, Humanoid character animation using AMP in 6 minutes, and cube rotation with Shadow Hand in 35 minutes all on a single NVIDIA A100 GPU. The platform also demonstrates sim-to-real transfer results on ANYmal and TriFinger, showcasing the ability of the simulator to perform high-fidelity contact-rich manipulation. The results show that Isaac Gym provides a high-performance training platform for robotics, with significant improvements in training speed and efficiency compared to conventional RL training methods.