NVIDIA has introduced Isaac Lab-Arena, an open source framework designed to simplify and scale robotic policy evaluation in simulation. Built as an extension to NVIDIA Isaac Lab and co-developed with Lightwheel, the tool helps developers quickly create, diversify, and run large-scale evaluations of generalist robot policies across different tasks, robots, and environments—without needing to build custom infrastructure from scratch. The pre-alpha release of Isaac Lab-Arena enables users to go from zero to a working evaluation setup with minimal effort. It offers streamlined APIs for task curation, automatic task diversification, and high-throughput, parallel evaluation across thousands of environments using GPU acceleration. This allows researchers and engineers to test policies at scale, even with complex, real-world-like scenarios. Key features include the ability to easily swap objects, robots, or scenes in a task without rewriting code. For example, a microwave-opening task can be quickly adapted to work with a power drill or a cracker box, or moved from a kitchen to an industrial setting. This flexibility supports rapid prototyping and benchmarking across diverse conditions. Isaac Lab-Arena also integrates seamlessly with data collection and training workflows. It works with tools like Isaac Lab-Teleop and Isaac Lab-Mimic, and supports post-training and inference of NVIDIA’s Isaac GR00T N models. This creates a closed-loop pipeline from training to evaluation. The framework is open source under a commercial license, allowing free use, distribution, and contribution. It can be deployed locally or in cloud environments like OSMO, and is already integrated into platforms such as the Hugging Face LeRobot Environment Hub. This enables developers to access and share environments, benchmarks, and evaluation methods across the robotics community. NVIDIA is collaborating with benchmark creators to bring existing evaluations onto Isaac Lab-Arena. Lightwheel has already contributed over 250 tasks through its RoboCasa and LIBERO task suites, and is developing RoboFinals, an industrial benchmark for complex real-world scenarios. RoboTwin is using the framework to expand its RoboTwin 2.0 simulation benchmark, while NVIDIA’s GEAR Lab and Seattle Robotics Lab are applying it to evaluate generalist robot models at scale. Future updates will include advanced capabilities like natural language-based object placement, composite tasking by chaining skills, reinforcement learning task setup, and heterogeneous parallel evaluations. Long-term goals involve using AI-driven simulation tools such as NVIDIA Cosmos and Omniverse NuRec to generate realistic, dynamic environments from real-world data. To get started, developers can explore the GitHub repository and documentation. The framework is designed to evolve with community input, so feedback is encouraged. For those new to robotics, NVIDIA offers free courses and resources to help build foundational skills. The full ecosystem aims to accelerate progress in embodied AI by making simulation-based evaluation faster, easier, and more accessible.