Data

Human motions paired with text descriptions for language-to-motion.

MAV simulator framework for Gazebo.

The classic open-source 3D robotics simulator, tightly integrated with ROS.

GPU-native MuJoCo in NVIDIA Warp for massively parallel learning.

Ten million 3D objects for populating simulators and training vision.

Hundreds of ready-made SDF models of robots, objects and buildings.

The benchmark that defined visual odometry and driving perception.

Dynamic Animation and Robotics Toolkit.

Multi-sensor ground robot sequences indoors and outdoors with ground truth.

Large-scale benchmark and baseline for general object grasping.

Unified human motion capture archive for training humanoid controllers.

Sensor, LiDAR and motion-forecasting datasets with HD maps.

Gazebo packages for simulating underwater vehicles and manipulators.

A simulated part-based interactive environment with articulated objects.

A browser-based hexapod kinematics simulator to design gaits and legs.

Isaac Gym environments for training multirotor control policies.

Simulated autonomous surface vessel competition environment.

33 hours of highway driving with camera, GNSS and CAN logs.

3,600 hours of egocentric human video used to pretrain manipulation models.

Generate thousands of demonstrations from a handful of human ones.

Export Onshape CAD assemblies directly to URDF, SDF and MJCF.

6,000 real scanned objects with meshes, videos and point clouds.

A large-scale in-the-wild robot manipulation dataset.

Photorealistic simulated SLAM sequences with hard weather and lighting.

60k+ trajectories of a low-cost WidowX arm across 24 environments.

Simulated planes, quadplanes, rovers and boats for ArduPilot.

15 million video frames from 7 robot platforms for large-scale visual learning.

100 hours of first-person kitchen video with action and object labels.

17.7 million simulated parallel-jaw grasps on 8,872 ShapeNet objects.

The standard benchmark objects for manipulation, ready for simulation.