NVIDIA Isaac Lab-Arena

An open-source NVIDIA extension to Isaac Lab for composing simulation benchmarks and evaluating robot policies in parallel.

Isaac Lab-Arena is an open-source framework built on top of Isaac Lab, developed by NVIDIA together with the simulation-data company Lightwheel and formally introduced on NVIDIA's blog in January 2026; it is still in alpha. Large-scale evaluation used to mean hand-writing a scene and a script for every single task. Isaac Lab-Arena instead splits an environment into three freely combinable pieces: the scene (which objects are placed where), the embodiment (the robot along with its sensors and controllers), and the task (what needs to be accomplished). It supports perturbation testing of lighting, cameras, and object properties, tracks progress through subtasks such as grasping and placing, and can run thousands of environments simultaneously on a GPU. It plugs into LeRobot's EnvHub, so it can evaluate models such as GR00T and π0.5, and it already ships adapted versions of RoboCasa, LIBERO, and RoboTwin 2.0.

ExampleIn a comparison NVIDIA published on its blog, running the same batch of evaluations in parallel with Isaac Lab-Arena took 0.76 hours, versus 34.9 hours running them one at a time.

Also called

Isaac Lab Arena, IsaacLab-Arena

Related

NVIDIA Isaac Lab · Simulation-Based Evaluation · Benchmark · LeRobot EnvHub · RoboFinals (Lightwheel industrial-grade simulation evaluation platform) · RoboCasa

Sources

NVIDIA Technical Blog: Simplify Generalist Robot Policy Evaluation in Simulation with NVIDIA Isaac Lab-Arena
GitHub: isaac-sim/IsaacLab-Arena

As of

2026-09