Dataset Registry

Aggregate dataset spanning 60+ constituent datasets, 22 robot platforms, and 527 manipulation skills. The largest open robotics dataset collection.

Large-scale in-the-wild manipulation dataset with 76K+ episodes across 564 scenes and 52 buildings. Multi-camera Franka teleoperation.

Large-scale WidowX tabletop manipulation dataset with 53K+ episodes across 24 environments and 13 skills.

Franka multi-task dataset with 12 skills across 38 kitchen tasks. Includes teleoperated, kinesthetic, and heuristic demonstrations.

Simulated bimanual cube transfer task with 50 human-collected episodes. Ideal for quick ALOHA experiments.

Simulated bimanual peg insertion task with 50 human-collected episodes.

Real bimanual tape dispensing task with 50 teleoperated episodes on static ALOHA.

Real mobile bimanual cabinet manipulation with 85 episodes on Mobile ALOHA.

Large-scale dual-arm humanoid manipulation dataset with 92K+ trajectories across 217 tasks and 87 skills.

Google Robot tabletop manipulation dataset with 87K+ episodes, used to train RT-1. 17 object categories.

Contact-rich manipulation dataset with 110K+ episodes across 4 robots and 147 skills. Includes force/torque and audio.

Simulated continual learning benchmark with 130 tasks across 4 suites (Spatial, Object, Goal, LIBERO-10).

Long-horizon furniture assembly benchmark with 5K+ episodes across 8 assembly tasks at 3 difficulty levels.

Simulated benchmark with 100 manipulation tasks in CoppeliaSim. Procedurally generated demonstrations.

Simulated Sawyer benchmark with 50 tabletop manipulation tasks for meta/multi-task learning.

Language-conditioned block pushing with 600K+ real episodes and open-vocabulary instructions.

Language-conditioned WidowX manipulation with 26K episodes across 100 tasks for zero-shot generalization.

Teleoperated play data on Kinova Jaco 2 with language annotations. 1K+ tabletop manipulation episodes.

Task-agnostic Franka play data with language annotations. 3.6K episodes of tabletop manipulation.

Multi-arm manipulation dataset from Stanford ILIAD lab, converted to RLDS format.

Simple 2D T-block pushing task with 206 episodes. The canonical LeRobot demo dataset.

100-episode subset of DROID in LeRobot format. Ideal for quick experiments and CI.

Real bimanual coffee preparation task with 50 teleoperated episodes on static ALOHA.

Large-scale simulated kitchen manipulation: ~100 human demos per task across 100+ atomic/composite tasks, expanded to ~100k+ trajectories with MimicGen in MuJoCo. A standard pretraining corpus behind GR00T.

GPU-parallelized manipulation simulator and benchmark with millions of demonstration frames across diverse task families (pick-place, stacking, articulated objects, dexterous).

Long-horizon, language-conditioned tabletop manipulation benchmark across four env splits (A/B/C/D); ~2M teleoperated trajectories with ~24k language-annotated.

Canonical imitation-learning / offline-RL manipulation benchmark: proficient- and mixed-human teleop demonstrations across sim and real tasks (Lift, Can, Square, Transport, Tool Hang).

Auto-generated manipulation demonstrations: ~48k trajectories across 12 tasks synthesized from ~200 human demos, with diverse object/scene configs and multiple arms.

Multi-embodiment real-world teleoperation dataset: ~107k trajectories, 479 tasks, 96 object classes across Franka single-arm, AgileX dual-arm, UR-5e, and the Tien Kung humanoid.

Large-scale real-world dual-arm humanoid manipulation: 1M+ trajectories, ~2,976 hours, 217 tasks and 87 skills with dexterous hands and mobile bases. Supersedes AgiBot World Alpha.

Scalable bimanual manipulation data generator and benchmark: 100k+ domain-randomized expert trajectories across 50 dual-arm tasks and five embodiments (incl. Aloha-AgileX).

In-the-wild human demonstrations collected with handheld grippers, yielding hardware-agnostic, deployable manipulation policies (incl. bimanual and dynamic tasks).