Robotics Datasets
Open robot learning datasets including Open X-Embodiment, DROID and AgiBot World — manipulation and navigation data for training policies.
Datasets
- FineART: FineART: 40,543 bimanual teleoperation episodes (1,718 h, 151 tasks) with 533,913 subtask labels, CC BY 4.0, plus the FineART-VLA model for LeRobot.
- Gen-HumanEgo: Gen-HumanEgo: 1,847.7 hours of egocentric human demonstrations from GenRobot.AI with 3D hand reconstruction, large-FOV depth and task annotations, CC-BY-SA-4.0.
- NVIDIA SimReady Assets: NVIDIA SimReady Assets: thousands of CC-BY-4.0 OpenUSD objects with physics for PhysX, MuJoCo and Newton, free on Hugging Face for robot simulation.
- Eidon Tracker POV: A CC-BY-4.0 dataset: 1,274 hours of first-person household video paired with 24 Hz orientation from seven body-worn IMUs.
- HiFi-UMI-2K: HiFi-UMI-2K: 2,000 hours of open robot-free bimanual manipulation data, six synchronised views, 3 mm accuracy, LeRobot v3 format, CC-BY-4.0.
- Robotic Origami Challenge: Fold Plane Demonstrations: Robotic Origami Challenge dataset: 682 LeRobot episodes of bimanual dexterous paper-airplane folding, with six video streams and 60 tactile channels. CC-BY-4.0.
- Galaxea Open-World Dataset: The Galaxea Open-World Dataset: hundreds of hours of real mobile-manipulation demos in authentic homes and offices, with the open G0 VLA models.
- AgiBot World: AgiBot World is a large-scale real-robot manipulation dataset: 1M+ trajectories from 100+ robots across 100+ real scenarios. IROS 2025 award finalist.
- M2DGR: M2DGR is a multi-modal ground-robot SLAM dataset: fisheye and infrared cameras, 32-beam lidar, IMU and GNSS across 36 sequences with ground truth.
- ABC-130k: ABC-130k: the largest open-source robot teleoperation dataset — 130,919 bimanual manipulation episodes (3,598 hours) on two-arm YAM stations, Apache-2.0.
- DROID: DROID is a large-scale robot manipulation dataset with 76,000 demonstration trajectories across 564 diverse real-world scenes and 350 hours of interaction data.
- Open X-Embodiment (OXE): Open X-Embodiment is the largest open robotics dataset with 1M+ episodes from 22 robot types, collected across 34 research labs spanning 527 skills.