Isaac Lab-Arena
NVIDIA's open framework for authoring robot benchmarks and testing policies at scale
License
Apache 2.0 (runs on Isaac Sim, which has proprietary parts)
Code: Apache-2.0, per LICENSE.md. GitHub shows no licence because the file opens with a Markdown heading. Required runtime: Isaac Sim, which includes components under NVIDIA's proprietary Omniverse licence, and its EULA must be accepted to run Arena. Models and benchmarks used with Arena, such as GR00T, π0.5 or third-party task suites, keep their own licences. Checked 2026-11-06 against LICENSE.md and the README.
Overview
Isaac Lab-Arena is an open-source framework from NVIDIA for building robot benchmarks in simulation and evaluating policies in them at scale. It extends Isaac Lab: an environment is composed from a scene, a robot embodiment and a task, so one task can be reused across robots, objects and settings without new code. One policy can be run across thousands of parallel environments on a GPU or spread over several machines, and the results show at which step of a task a policy fails and which conditions it is sensitive to. It is alpha software, with 0.3.1 as the current release line, and published benchmarks such as RoboDojo, RoboTwin 2.0 and Lightwheel's RoboCasa and LIBERO task suites already run on it.
Isaac Lab-Arena is NVIDIA's framework for building simulation benchmarks and evaluating robot policies in them, written as an extension of Isaac Lab on Isaac Sim. Instead of a separate configuration for every combination of robot, object and setting, you compose an environment from three independent parts: a scene with its objects and furniture, an embodiment (the robot with its sensors, observations and actions) and a task (what the robot must do and how success is judged). Arena assembles them into a standard Isaac Lab environment, so moving a task from a Franka arm to a humanoid, or from a kitchen to a packing table, means changing one building block. Objects are placed with relations such as on and next to instead of coordinates, short skills chain into long tasks, and a variation system turns one environment into a sweep over lighting, backgrounds, camera parameters and object mass, a controlled form of domain randomization. An experimental agent turns a plain-language task description into an editable environment spec built from SimReady assets.
Evaluation runs one policy across thousands of environments on a GPU, and those environments can hold different objects. Experiments that combine several tasks and policies run locally or across nodes with OSMO. Policies sit behind a server and talk to Arena through one observation-and-action interface: the repository has integrations for GR00T, OpenPI's π0.5 and DreamZero, and other models plug in the same way. Results come back as aggregate and per-episode metrics. Subtask predicates record how far each rollout got (grasp, lift, transport, place), and sensitivity analysis shows which conditions a policy breaks under. Newton can replace PhysX as the physics backend for deformable and contact-rich tasks such as gear insertion.
In NVIDIA's launch post, evaluating GR00T N1.5 on 10 RoboCasa tasks with 4,096 environment variations per task took 0.76 hours in parallel on 8 GPUs, against 34.9 hours run one after another on the same framework, about 40 times faster. A preliminary test in the documentation measured 2,390 environment steps per second with 1,024 parallel environments on one RTX 5880 Ada GPU, without cameras.
Benchmarks live in their authors' own repositories and run on Arena. The README lists Lightwheel's task suites adapted from RoboCasa and LIBERO, RoboTwin 2.0, Isaac for Healthcare's RHEO workflows, VABAR, DexBench and RoboDojo. NVIDIA announced RoboDojo on Arena with bimanual tasks such as precision insertion, folding and tool use; RoboDojo's Arena branch runs 54 tasks on ARX's X5 dual-arm robot. Arena itself ships a Kitchen Benchmark of 31 DROID manipulation tasks and tasks modelled on RoboLab, listed in its example environments. Arena environments can also be shared on the LeRobot Environment Hub and evaluated with LeRobot's own evaluation command, as the Hugging Face post shows with SmolVLA.
Why it matters
Generalist policies such as vision-language-action models are judged by how they hold up across many tasks, robots and conditions, and testing that on real hardware is slow, costly and hard to repeat. Simulation is the cheap first check, but benchmarks have usually shipped their own environments, scripts and runners, so testing one policy on several of them means several separate setups. Arena puts them on one core: benchmark authors reuse the same scenes, robots and metrics, and a policy team can run all of them through one client. Its predicate and sensitivity tools also answer what a single success rate hides: at which step a policy failed and under which conditions. Our catalogue has other evaluation tools, SimplerEnv and Inspect Robots. Arena is the one built for authoring new benchmarks and running them at GPU scale, and it is the base that RoboDojo, RoboTwin and Lightwheel build on and that LeRobot's Environment Hub supports.
Problem it solves
Evaluating a generalist robot policy properly means running it on many tasks, robots, objects and conditions. In simulation that usually means a hand-written environment and runner for every combination, with rollouts run one after another. Arena replaces that with reusable building blocks, parallel evaluation on the GPU and shared metrics, so a benchmark is written once, varied by configuration and run at scale, and different benchmarks can be run through the same policy interface.
Key takeaways
- Environments are composed from separate scene, embodiment and task blocks and build into standard Isaac Lab environments
- Runs one policy across thousands of parallel environments on a GPU, or across machines with OSMO
- Policies run behind a server with one interface, with integrations for GR00T, π0.5 and DreamZero included
- Subtask progress and sensitivity analysis show where a policy fails and under which conditions
- Published benchmarks built on it include RoboDojo, RoboTwin 2.0 and Lightwheel's RoboCasa and LIBERO task suites
- Alpha software: APIs change between releases and NVIDIA says not to use it in production
Who it is for
- Teams training vision-language-action or other generalist policies who need repeatable evaluation across many tasks
- Benchmark authors who want to publish tasks that others can run without rebuilding scenes, robots and metrics
- Isaac Lab users who want evaluation, controlled variations and failure analysis on top of their environments
- LeRobot users who want to share simulation environments through the Environment Hub
Limitations
- Alpha software (version 0.3): public APIs change without deprecation warnings, and NVIDIA advises against production use
- Linux only (Ubuntu 22.04 or later), with an NVIDIA GPU that meets Isaac Sim's hardware requirements
- Runs on Isaac Sim, which includes components under NVIDIA's proprietary licence terms, so the full stack is not open source
- Tied to NVIDIA's stack: every environment is an Isaac Lab environment, so benchmarks built on Arena run only there
- Versions are release branches rather than tagged GitHub releases, and each pins specific Isaac Lab and Isaac Sim versions
- The agentic environment generation is marked experimental, and its agents and design may change