Browse a curated catalog of datasets for humanoid robots and embodied AI, spanning real-world demonstrations, teleoperation, motion capture, egocentric vision, simulation, manipulation, locomotion, and cross-embodiment robot learning.
90 Datasets · Page 2 of 8
Ego4D contains more than 3,670 hours of first-person daily-life video from 923 participants across 74 locations in nine countries. Portions include audio, narrations, gaze, stereo video, synchronized wearable views, and 3D scans, with benchmarks for episodic memory, hand-object interaction, social understanding, diarization, and forecasting.
Ego-Exo4D captures 1,286.3 hours of synchronized first- and third-person video across 5,035 takes, 740 participants, and eight skilled-activity domains. Aria glasses and multiple external cameras provide calibrated video, audio, eye gaze, inertial data, trajectories, point clouds, and three forms of language annotation for skill and cross-view understanding.
DexYCB contains 582,000 RGB-D frames from 1,000 sequences of 10 people grasping 20 YCB objects, recorded synchronously from eight viewpoints. It provides calibrated imagery and 3D hand and object poses for hand-pose estimation, 6D object-pose estimation, and safe human-to-robot handover research.
robomimic combines proficient-human, multi-human, and machine-generated demonstrations for five simulated and three real robot manipulation tasks. The HDF5 datasets include low-dimensional or multi-camera image observations for pick-and-place, insertion, tool hanging, and bimanual transport, supporting reproducible imitation and offline reinforcement learning.
CALVIN contains 24 hours of teleoperated play data, with six hours in each of four simulated tabletop environments and roughly 20,000 language directives. The benchmark evaluates language-conditioned policies on long sequences of object manipulation using static and gripper cameras, proprioception, depth, or tactile observations.
ManiSkill2 provides more than four million demonstration frames for 20 simulated manipulation skill families and over 2,000 object models. Its SAPIEN environments span stationary and mobile platforms, single- and dual-arm control, rigid and soft bodies, RGB-D and point-cloud observations, and tasks from picking and insertion to pouring and excavation.
ALOHA Static comprises 825 real demonstrations of fine-grained bimanual tabletop manipulation across 12 tasks, including sealing a bag, wrapping candy, opening a cup, preparing tape, using a coffee machine, fastening cable, slotting a battery, and handing over tools. The fixed ALOHA workcell records 14-dimensional joint state and action data at 50 Hz alongside four 480×640 RGB views from two stationary and two wrist-mounted cameras. The task-specific releases are distributed in LeRobot format for imitation-learning research and co-training.
RoboArena is a distributed real-world evaluation dataset for generalist policies on the DROID robot platform. The July 17, 2026 snapshot contains 3,883 evaluation sessions and 10,783 autonomous policy episodes, with 27,148 multi-view videos, matching proprioception and action files, task instructions, success scores, pairwise preferences, and evaluator feedback. Double-blind comparisons let participating institutions choose their own tasks and environments while preserving matched conditions within each policy comparison.
AIRoA MoMa v1.1 provides 23,762 filtered Human Support Robot teleoperation episodes totaling 87 hours across seven household task families. The gated LeRobot release pairs head- and hand-camera RGB video with robot state, calibration, task-success metadata, and hierarchical short-horizon and primitive-action annotations; the research dataset also records synchronized wrist force-torque signals. Nineteen operators used eight HSR robots to collect tasks including towel handling, coffee making, dishwashing, toast preparation, lamp control, and slipper organization.
MolmoBot-Data contains 1.7 million procedurally generated expert trajectories totaling 5,704 hours across eight manipulation task types on Franka and RB-Y1 platforms. MolmoBot-Engine uses task-and-motion planning and randomizes objects, receptacles, lighting, and camera poses across more than 94,000 simulated indoor environments. The data supports articulated-object interaction, pick-and-place, and mobile manipulation research, including zero-shot sim-to-real policy training.
MolmoAct2 Bimanual YAM is a large-scale collection of real bimanual manipulation demonstrations gathered on the custom Yet Another Manipulator platform. The merged LeRobot release contains 32,246 episodes, 76 million frames, and 34 language-annotated tasks with top, left, and right RGB views plus 14-dimensional joint state and action data; the full collection is described as more than 720 hours of training data. Tasks range from folding clothes and untangling cables to scanning groceries, packing medication, and bussing tables.
MolmoAct2 SO-100/SO-101 curates 38,059 demonstrations, 19.8 million frames, and about 184 hours from more than 1,200 public LeRobot repositories contributed by 377 users. A four-stage pipeline checks structural validity, excludes evaluation data, enforces source eligibility, and applies a TOPReward quality gate while retaining varied tasks, camera setups, objects, environments, and both low-cost arm embodiments. The release also supplies per-episode language annotations for the source datasets.