IsaacLab
Unified framework for robot learning built on NVIDIA Isaac Sim
Isaac Lab is an open-source Python framework for training robots using reinforcement learning, running thousands of GPU-parallel simulations on NVIDIA Isaac Sim to dramatically speed up robotics research and sim-to-real transfer.
Isaac Lab is an open-source Python framework for robotics research, built on top of NVIDIA Isaac Sim. Its main purpose is to give researchers and engineers a single, consistent place to train robot software using techniques like reinforcement learning, imitation learning, and motion planning. Reinforcement learning is an approach where a robot (or a simulated version of one) learns through trial and error by receiving feedback on whether its actions were good or bad.
A key feature is that simulations run on the GPU rather than the CPU, which means thousands of virtual robot experiments can happen in parallel, drastically reducing the time needed to train a robot to perform a task. The framework also includes realistic sensor simulation, covering cameras (with depth and segmentation), LIDAR, inertial measurement units, and contact sensors. This matters for research aimed at sim-to-real transfer, where behavior learned in a virtual environment is transferred onto a physical robot.
The library ships with more than 16 robot models spanning different categories: robot arms, quadruped robots (four-legged), and humanoids. It also comes with over 30 pre-built training environments that are ready to use with several popular reinforcement learning libraries. Simulated physics covers rigid objects, articulated systems (jointed structures like a robot arm), and deformable objects. Simulations can run on a single local machine or be distributed across cloud infrastructure.
Isaac Lab requires NVIDIA Isaac Sim to run, and Isaac Sim itself is licensed under proprietary terms from NVIDIA, which is a notable constraint for anyone considering this for commercial use. The core Isaac Lab framework is released under the BSD-3 license, with one extension (isaaclab_mimic) under Apache 2.0. A technical paper describing the system in detail is available on arXiv, and the project includes documentation, tutorials, and a community discussion space on GitHub.
Where it fits
- Train a robot arm to pick and place objects using reinforcement learning in simulation without needing physical hardware.
- Run thousands of parallel robot training experiments on a GPU to find an effective motion policy in hours instead of weeks.
- Transfer a walking policy trained in Isaac Lab simulation onto a physical quadruped robot.
- Benchmark a new reinforcement learning algorithm using the 30+ pre-built training environments included with the framework.