nvlabs/mobilitygen

MobilityGen helps you generate training data for mobile robots, things like wheeled robots, quadrupeds, and humanoids, without needing a physical robot or a real-world testing environment. Instead of driving an actual robot around a warehouse to collect camera images, depth readings, and position data, you do it all in a realistic 3D simulation. The result is datasets you can use to train navigation models or test robotics algorithms. The tool runs on top of NVIDIA Isaac Sim, which is a physics-accurate virtual environment for robotics. You pick a scene (like a warehouse), choose a robot (such as the Boston Dynamics Spot or a Unitree H1 humanoid), and then decide how you want to drive it around. You can steer manually with a keyboard or gamepad, or you can let the system drive automatically using random movements or path-following. As the robot moves, the system records everything, RGB camera images, depth maps, segmentation images, robot poses, joint positions, and occupancy maps, into structured files you can load into your machine learning pipeline. This is useful for robotics researchers, engineers, or teams building autonomous navigation systems who need large amounts of labeled data. Real-world data collection is slow, expensive, and hard to annotate. With a simulated environment, you can generate thousands of training examples with perfectly accurate labels (ground truth) because the simulation knows exactly where everything is. A team training a robot to navigate a warehouse, for example, could generate varied scenarios with different layouts and movement patterns without ever setting foot on a warehouse floor. The project is built to be extensible. If you have a robot that isn't included by default, you can add it by creating a new robot class that defines how it moves and where its sensors are. The same goes for movement scenarios, if the built-in random or path-following options don't fit your needs, you can define your own logic for how the robot explores the scene. There's also an optional GPU-accelerated path planner for better performance on NVIDIA hardware, though the README notes this requires specific CUDA setup and recent GPU architecture.