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 3 of 8
MolmoAct Dataset contains 10,689 high-quality trajectories of a single-arm Franka robot performing 93 manipulation tasks in home and tabletop environments. Five full-time operators gathered the demonstrations over two months, with the data released in LeRobot format. The release supports mid-training action-reasoning models that connect instructions and visual observations to spatial plans and low-level robot actions.
FTP-1 Dataset aggregates about 3,000 hours of tactile manipulation data from 26 human and robot sources across 21 sensors. It spans image-based, array-based, and state-based touch on dexterous hands and gripper robots, with tactile annotations normalized through the Morphology-Aware Tactile Token Space and instructions rewritten for linguistic diversity. The mixture is designed to pretrain transferable contact-rich manipulation policies across sensors and embodiments.
The LESS release contains more than 800 hours of tactile palpation measurements for reconstructing the internal structure of soft breast phantoms. A Franka Panda with a gel-based tactile sensor collected controlled poke and motion-primitive trajectories, while additional sets cover larger phantoms, multiple inclusions, and hand-held sensing. Force and pose sequences are paired with MRI-derived geometry for training and evaluating local 2D and 3D tactile imaging models.
The LeRobot shirt-folding release contains 5,688 teleoperated demonstrations, about 131 hours, collected across eight bimanual OpenArm setups. Each episode records 16-dimensional joint-and-gripper states and actions at 30 fps with one base and two wrist cameras in LeRobotDataset v3.0 format. A curated 1,200-episode, roughly 30-hour subset combines the best recordings with additional demonstrations collected under a more consistent folding strategy.
RoboDojo is an evaluation-focused benchmark spanning 42 Isaac Sim tasks and 18 real-world manipulation tasks across three robot embodiments. Its tasks probe generalization, memory, precision, long-horizon execution, and open-vocabulary instruction following, while the release provides benchmark assets, configuration validation, seed-controlled layouts, and result artifacts. RoboDojo-RealEval standardizes hardware, scene resets, evaluation protocols, and deployment interfaces for reproducible physical testing.
ArmnetBench v0.1 contains 3,718 labelled real-robot episodes across eight single-arm and four bimanual SO-101 tasks. It combines 600 human-teleoperated reference demonstrations with rollouts from seven learned policies, recording three camera views at 20 fps alongside state, action, task, provenance, and success-quality labels. The release is available in LeRobot v3.0 and RoboMeter formats for policy evaluation, reward modelling, and downstream training.
TrajBooster's approximately 35-hour, 30 GB dataset contains 1,960 episodes across 176 AgiBot-World tasks retargeted from a wheeled humanoid to Unitree G1. It extracts dual-arm 6D end-effector trajectories from real source data, tracks them with a whole-body controller in Isaac Gym, and replaces source actuator commands with G1-compatible actions while preserving source vision and language, producing heterogeneous vision-language-action triplets. The data supports cross-embodiment VLA post-pre-training before adaptation with about ten minutes of target-robot teleoperation.
Arena G1 Loco-Manipulation contains five human-teleoperated seeds and 50 automatically generated Unitree G1 trajectories for a simulated box pick-and-place task requiring navigation in Isaac Lab. The MimicGen episodes run at 50 Hz and include 26-DoF desired joint actions and states, timestamps, task-description indices, validity annotations, and 256×256 first-person RGB video. It is intended for behavior cloning, generalist policy post-training, and sim-to-sim or sim-to-real loco-manipulation research.
G1 Moves provides 60 human-performance clips totaling 29.6 minutes, primarily dance and karate, for the 29-DoF Unitree G1. Fifty-nine clips were acquired with markerless LiDAR-and-vision motion capture and one from monocular video, then retargeted into G1 joint trajectories. Each clip includes raw BVH and FBX motion, root and joint trajectories, MuJoCo-derived training states and velocities, validation metadata, and deployable ONNX imitation policies for retargeting, reinforcement learning, visualization, and sim-to-real study.
OmniRetarget generated more than nine hours of interaction-preserving humanoid trajectories from OMOMO, LAFAN1, and in-house motion capture; the public Unitree G1 release contains four hours because LAFAN1-derived data cannot be redistributed. It covers object carrying, terrain climbing and traversal, and combined object-terrain interaction, with systematic augmentation of object pose, object size, terrain, and embodiment. NPZ files store frame rate, 29-DoF robot joint positions, floating-base pose, and optional object pose for loco-manipulation and reinforcement-learning research.
MOSAIC uses about 64 hours of heterogeneous motion data: 3.1 hours of optical motion capture, 7 hours of inertial capture, 51 hours of public corpora, 2.2 hours of curated GENMO motion, and 1 hour of interface-adaptation data. The release includes AMASS-style human motion, Unitree G1-retargeted NPZ trajectories, raw inertial streams, generation prompts, and roughly 30 minutes each of PICO VR and Noitom data. It supports generalist motion tracking, offline replay, and robust whole-body teleoperation.
ALMI-X contains approximately 81,500 four-second, 200-step Unitree H1-2 trajectories generated by ALMI policies in MuJoCo. It combines AMASS-derived upper-body motions with omnidirectional lower-body velocity commands, including standing, turning, and different movement speeds, and assigns a templated language description to each combination. Files provide robot observations, actions, joint positions, global position, and global orientation for training language-conditioned humanoid locomotion and whole-body control models deployable on real robots.