Search public pages, research tools, and SAVRN solutions.
SAVRN Model Hub · Datasets by Task
95 open-weight robotics datasets in the SAVRN Model Hub, with National Engineering Research Center for Big Data Software, IPEC at Shanghai AI Laboratory and K publishing the most.
Robotics data is the scarce input of physical AI, and this page shows who is filling the gap. NVIDIA's GR00T simulation trajectories lead in downloads, followed by Genrobot's collection of more than 13,000 hours of real two-handed manipulation, and a long list of classic datasets such as DROID, Language Table and KUKA converted to the LeRobot format.
Most of these sets are Apache 2.0 or CC-BY-4.0, which makes them usable for commercial training with attribution. The practical constraint is size: real-world robot data runs to hundreds of gigabytes or more, so plan storage and bandwidth before you plan training runs. Mixing public data with demonstrations from your own robot remains the most reliable recipe.
| Dataset | Publisher | License | Monthly downloads |
|---|---|---|---|
| PhysicalAI-Robotics-GR00T-X-Embodiment-Sim | NVIDIA | cc-by-4.0 | 1.3M |
| 10Kh-RealOmin-OpenData | Genrobot.ai | cc-by-sa-4.0 | 689.4k |
| language_table_lerobot | IPEC at Shanghai AI Laboratory | apache-2.0 | 474.3k |
| droid_lerobot | IPEC at Shanghai AI Laboratory | apache-2.0 | 361.9k |
| droid_1.0.1 | Remi Cadene | apache-2.0 | 352.1k |
| kuka_lerobot | IPEC at Shanghai AI Laboratory | apache-2.0 | 316.5k |
| ABC-130k | XDOF | apache-2.0 | 296.4k |
| AgiBotWorld2026 | AgiBot World | cc-by-nc-sa-4.0 | 245.9k |
| stereo-550 | FPV Labs | other | 242.8k |
| ACE-Data-0 | ACE Robotics | other | 242.5k |
| License | Datasets | Commercial use |
|---|---|---|
| apache-2.0 | 57 | Yes |
| other | 10 | Read the license |
| cc-by-4.0 | 8 | Yes |
| not stated | 7 | Not stated |
| cc-by-nc-4.0 | 6 | Not without separate permission |
| mit | 2 | Yes |
| cc-by-sa-4.0 | 2 | Yes |
| unknown | 1 | Read the license |
| Publisher | Datasets |
|---|---|
| National Engineering Research Center for Big Data Software | 12 |
| IPEC at Shanghai AI Laboratory | 5 |
| K | 4 |
| Kun Zhao | 3 |
| Anuj Zore | 2 |
| VR-VLA | 2 |
We provide a set of datasets used for post-training of GR00T N1. Each dataset is a collection of trajectories from different robot embodiments and tasks. Users can download a specific subset of data by specifying the dataset name. 1. Option 1: with huggingface-cli Replace gr1armsonly.CanSort/ with the dataset name you want to download. 2. Option 2: Github LFS Replace the "gr1armswaist.CupToDrawer" with the intended dataset folder
Boasting over 13,000 hours of cumulative data and 5 million+ clips, it ranks as the largest open-source embodied intelligence dataset in the industry. Update Notes：Stage 3 data upload completed. 13,000+ hours of pure dual-hand data with frame-level alignment latency < 1ms Full high-precision trajectory reconstruction, breaking the limit of superficial open source, fully ready-to-use 3,000+ contributors and 10,000+ real household scenarios with exceptional diversity
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Real-World Embodied Intelligence Dataset As robotics research advances into real-world scenarios, the demand for authentic, high-quality data has become increasingly urgent. Following AGIBOT WORLD's "ImageNet moment," we now release the AGIBOT WORLD 2026 dataset. Built upon massive real-world scenes, it systematically spans pivotal research directions in embodied intelligence, designed to power the next generation of embodied agents. The AGIBOT WORLD 2026 dataset is collected from 100% real-world environments, covering commercial spaces, home, and other general-purpose scenarios. Collected on the AGIBOT G2 robot platform through a free-form collection mode, the dataset provides developers…
FPV Labs Open-Source Stereo Hardware A first-person calibrated stereo RGB video dataset capturing everyday human manipulation across objects, materials, tools, and multi-step activities. Every session is recorded as a synchronized left/right camera pair with per-session stereo calibration, giving the visual geometry of hands, object interaction, state
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot. One of the biggest open-source dataset for robotics with 27.044,326 frames, 92,223 episodes, 31,308 unique task description in natural language. Ported from Tensorflow Dataset format (2TB) to LeRobotDataset format (400GB) with the help from IPEC-COMMUNITY.
This dataset was created using LeRobot.
Retargeted AMASS for Robotics Project Overview This project aims to retarget motion data from the AMASS dataset to various robot models and open-source the retargeted data to facilitate research and applications in robotics and human-robot interaction. AMASS (Archive of Motion Capture as Surface Shapes) is a high-quality human motion capture dataset, and the SMPL-X model is a powerful tool for generating realistic human motion data. By adapting the motion data from AMASS
Retargeted AMASS for Robotics Project Overview This project aims to retarget motion data from the AMASS dataset to various robot models and open-source the retargeted data to facilitate research and applications in robotics and human-robot interaction. AMASS (Archive of Motion Capture as Surface Shapes) is a high-quality human motion capture dataset, and the SMPL-X model is a powerful tool for generating realistic human motion data. By adapting the motion data from AMASS
This project aims to retarget motion data from the AMASS dataset to various robot models and open-source the retargeted data to facilitate research and applications in robotics and human-robot interaction. AMASS (Archive of Motion Capture as Surface Shapes) is a high-quality human motion capture dataset, and the SMPL-X model is a powerful tool for generating realistic human motion data. By adapting the motion data from AMASS to different robot models, we hope to provide a more diverse and accessible motion dataset for robot training and human-robot interaction. This open-source project includes the following: 1. Retargeted Motions: Motion files retargeted from AMASS to various robot models.…
We introduce Hy-Embodied-0.5-VLA (Hy-VLA) — an end-to-end Vision-Language-Action system that spans the full robot learning stack: data collection, model design, pre-training, supervised fine-tuning, RL post-training, and real-world deployment. Built on the Hy-Embodied-0.5 MoT backbone, Hy-VLA integrates a flow-matching action expert, a compact memory encoder for multi-frame history, and a delta-chunk action representation decoupled from embodiment-specific kinematics. Powered by 10,000+ hours of high-fidelity UMI demonstrations collected via a custom fingertip interface with optical motion-capture, Hy-VLA achieves state-of-the-art results on the RoboTwin 2.0 benchmark (90.9% / 90.1% on…
First-person stereo video, depth and head pose of real shelf-work and facility tasks, captured on a Pico VR headset. By OpenElephant Intelligence (公象智能), Beijing. This is the free, open entry point to a much larger commercial collection (10,000+ hours). It is meant for evaluating the data quality and format before requesting the Advanced Edition or a commercial license. Real store shelving with beverage coolers, snack gondolas and hanging snack racks. Most tasks are "error correction": the shelf starts out mis-stocked and the operator finds and fixes the misplaced items. A large indoor sports hall with fitness equipment, mats and event cabling. spatialcalibration.json gives, for the left…
A massive, highly-diverse dataset of Fused Deposition Modeling (FDM) G-codes generated directly from Printables. This dataset is specifically designed for training machine learning models on raw 3D printing manufacturing instructions (G-code). It can be used for tasks like G-code generation, print failure prediction, semantic analysis of toolpaths, and printer-agnostic slice classification. To ensure a highly robust and diverse set of training data, every original 3D model from the source dataset has been iteratively sliced into multiple variants (default: 10 per model) using PrusaSlicer. The pipeline supports and natively slices.stl,.obj,.3mf,.step, and.amf files. For each variant, a…
First-person stereo video, depth, head pose, pose-aligned point clouds and time-stamped action captions of real shelf-work and facility tasks, captured on a Pico VR headset. By OpenElephant Intelligence (公象智能), Beijing. The Advanced Edition adds 3D geometry and language to the Basic Edition: every episode comes with world-aligned point clouds, a ready-to-view MCAP recording, and a sequence of time-stamped captions describing what the operator is doing. It is a sample of a 10,000+ hour commercial collection. Gated. Requests are reviewed manually, usually within 3 business days. Real store shelving with beverage coolers, snack gondolas and hanging snack racks. Most tasks are "error…
ABC-130k ABC-130k is the largest open-source robot teleoperation dataset. It contains bimanual manipulation trajectories collected on two-arm YAM stations. Episodes are distributed as MCAP files, with subtask annotations kept as separate artifacts so they can be revised or extended independently of the underlying episode data. For details on the accompanying paper, see abc.bot. Please see the GitHub repo here for code to train and deploy with this dataset. Dataset
ACE-Data-0 Human-Centric Ambient Capture as Embodied Data Engine S-Lab, Nanyang Technological University, Singapore ACE Robotics ACE turns real home environments into spatially calibrated, temporally synchronized recording studios for embodied AI. ▶ Demo video Full story, figures, and interactive examples on the blog What this is Learning to act in the physical
HiFi-UMI-2K: High-Fidelity Robot-Free Manipulation Data 2,000 hours released · 6 accuracy · <40 µs synchronization Project Website | Dataset | Paper: arXiv:2607.25895 Examples from the HiFi-UMI corpus. Click the image to play the video. Introduction HiFi-UMI is a portable, high-fidelity bimanual capture system for collecting robot-free manipulation demonstrations.
Gen-HumanEgo 1,800+ hours of egocentric human demonstrations with synchronized, structured supervision for embodied AI and robot learning. Gen-HumanEgo contains real-world first-person demonstrations collected across diverse tasks, people, environments, and ways of performing activities using a unified six-camera DAS-Ego setup. GenRobot's Data Foundation Model (DFM) processes the recordings to provide complementary training signals for human motion, spatial understanding
The D1 dataset is exported from the D1 headset multi-camera human motion capture system developed by Delta Intelligence. Each recorded episode contains synchronized multi-view video streams and whole-body skeleton and headset pose data. The dataset supports research and development in: - Human detection, keypoint detection, and pose estimation - Egocentric visual observation and synchronization checking - Imitation learning, behavior cloning, and embodied intelligence model training - Visual-inertial odometry and SLAM-related observation analysis Unlike purely third-person datasets, D1 combines egocentric RGB and fisheye views with global whole-body skeleton and headset pose states, making…
Accepted zero-latency teacher rollouts for the supported benchmark tasks. - humanoidbenchbalancesimple: 90 training and 10 validation episodes. The observation.image values are PNG bytes declared as the Hugging Face Image feature, so the Dataset Viewer renders them instead of showing their encoded representation. Canonical LeRobot MP4 files remain under each split's videos/ directory. - mikasainterceptgrabfast: fixed 250/25 StarVLA split. Every Viewer row includes its corresponding observation.images.top and observation.images.wrist frame. The canonical LeRobot data remains under the task path with its RGB observations stored in MP4 files.…
Per-frame metric scene depth (Depth-Anything-V2 Metric Indoor Large, 20 m ceiling) for the episodes inpainted in VR-VLA/VR-egodex-inpaint-cache-full. Depth is the second input Eq. 4 compositing needs beside the inpainted plate; together they let scripts/egodex/buildtask.py retarget an episode onto all 16 embodiments. Archive names are IDENTICAL to the inpaint repo, one tar per /: Episode numbers match the source dataset's episode%06d.mp4. is part of the key: most task names exist in both train/ and test/ with overlapping episode numbers. Archives are uncompressed tar (PNG payload). train/addremovelidv2.1.tar (308 GB, 1,158 episodes) exceeds the Hub's 50 GB file limit and is stored as…
English | 中文 Physically-annotated indoor scene USD assets derived from 14,169 real interior design schemes on the sunvage floorplan platform — every furniture instance carries collision shapes, mass, friction and restitution, and opens directly in USD physics simulators. sunvage-vertens3d-simhome is a collection of physically-annotated indoor scene USD assets derived from real interior design schemes in the sunvage (三维家) floorplan library. It provides both whole-home multi-space scenes and individual single-space scenes. Every physical instance carries collision shapes and mass properties, and the scenes can be loaded directly into USD physics simulators such as Isaac Sim / Isaac Lab for…
Teleoperated bimanual manipulation demonstrations collected in MuJoCo with a BimanualYAM robot, packaged in LeRobot v3.0 format for pi0.5 training. Three RGB views per frame — camerafront, cameraleft, cameraright (224x224, H.264). Robot state and action are both 14-dimensional: 6 arm joints plus 1 gripper per arm. A 128-dimensional non-robot state carries object/scene information for diagnostics. Actions are stored as absolute joint targets. The training recipe converts the 12 arm dimensions to deltas relative to the current action-chunk anchor while the 2 gripper dimensions stay absolute. The normalization statistics in meta/stats.json are computed in that 50-step relative space — do not…
DexCanvas: Bridging Human Demonstrations and Robot Learning for Dexterous Manipulation DexCanvas is a dataset of 111,857 complete right-hand trajectories and 55,442,683 frames of rigid hand–object interaction. Every Parquet row is one trajectory with a shared time axis for hand kinematics, object pose, and per-frame contacts. For each trajectory, the repository also provides the complete dual-camera JPEG stream of the corresponding capture and shape-specific kinematic assets for replay. Camera images are associated at the complete Raw trajectory level. The release does not claim that trajectory frame i corresponds to image i. - 111,857 trajectories; 55,442,683 frames; 96,484,004 contact…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
1,274 hours of egocentric video paired with 7-point IMU arm tracking, recorded during ordinary household work. Contributors wore a head-mounted camera and a seven-sensor IMU harness while doing real chores in their own homes: laundry, cleaning, dishes, cooking. Each recording pairs first-person video with 24 Hz orientation data for both hands, both forearms, both upper arms, and the chest. This is a complete, final release. Eidon AI (Solidic Labs Inc) has wound down; the dataset is published under CC-BY-4.0 so it outlives the company. The release is split across two repos and a bucket. The data/NNN/ split exists only to keep directories manageable; the grouping carries no meaning. Both are…
This dataset was created using LeRobot.
EgoDex Retarget — 16 robots EgoDex egocentric human demonstrations retargeted onto 16 robot embodiments. For every episode, the human wrist and hand motion is solved into a joint trajectory for each robot, and the robot is rendered and composited into the original scene, over hand-erased frames and per-frame metric depth. Each (robot, task) pair is one LeRobot v3.0 dataset. VR-VLA/VR-egodex-inpaint-cache-full (hand-erased frames), VR-VLA/VR-egodex-depth-cache-full
GR00T-N1.7 LIBERO-X backbone features — 90-task fine-tune (LEVEL1-3) Aligned rollouts of a LIBERO-X fine-tune of GR00T-N1.7 (rohansiva/gr00t-libero-x-90task) on the LIBERO-X simulator, over the exact 90 tasks that checkpoint was fine-tuned on (30 tasks × 3 difficulty levels, LEVEL1–LEVEL3). Train and eval task sets are identical by design, so this is an in-distribution dataset for the checkpoint. 90 tasks × 20 rollouts = 1,800 episodes (600 per level). Every GR00T inference
SlugTrails is an egocentric floor-plan localization benchmark for large indoor buildings. This Hugging Face package (slugtrails-1.0) is the canonical public data blob: anonymized RGB, Project Aria MPS trajectories, floor-plan–aligned ground truth, CAD-derived maps, and official evaluation / finetune splits. manifests/ holds our official splits for the paper protocol (SHA-256 digests in RELEASE.json). You may define your own sampling if you prefer; numbers comparable to the paper should use these manifests. - 3 campus buildings / 6 floors — floor-plan outline 22 089 m², circulation space 5317 m² (same figures as the paper) - 6 Aria sessions (~30 Hz RGB); trajectory length ≈ 1.80 km (≈…
This dataset was created using LeRobot.
This sample of Pantheon DiverseUMI contains 25 hours of bimanual tabletop manipulation recorded with two handheld UMI grippers and an exocentric top camera. Each episode is a single task, such as stacking two blocks or passing a ball between grippers. Every episode includes three synchronized videos (top camera and both wrist cameras), per-frame end-effector actions for both wrists, and dense annotations of what the operator does. The sample is split into two halves of roughly equal size: something over, and then recover. Tasks range from 1 second to about 10 minutes, with a median of 7 seconds. Of the 4,588 tasks, 4,391 succeed, 154 fail and 43 are partially completed. For access to more…
Apache TsFile version of weiye11/0603pi0onlyright. A right-arm-only teleoperation dataset recorded on an SO-100 robot and packaged in the LeRobot v2.1 layout. Each row is one control step: a 6-dimensional right-arm joint/gripper state and the 6-dimensional commanded action. The source dataset card does not name the manipulation task; the metadata records a single task (taskindex = 0). - Time (INT64, milliseconds) — round(timestamp 1000); restarts at 0 for each episode. - episodeindex (TAG) — episode / device dimension. - taskindex (TAG) — task dimension (constant 0 here). - frameindex (FIELD, INT64) — source frame counter; sampleindex (FIELD, INT64) — source index column (renamed).…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Apache TsFile version of nitsuga-hug/augs-hiliter01. A SO follower (SO-101/SO-100) dataset for the "pnphiliter" pick-and-place task, collected with augmented human-in-the-loop data. Each frame records the 6-D joint state and the 6-D commanded action. Task (taskindex 0): "pnphiliter". - Time (INT64, milliseconds) - per-episode sample time, round(timestamp 1000) in milliseconds (the source timestamp column equals Time / 1000 seconds and is therefore not stored separately). - episodeindex, taskindex (TAG) - device dimensions identifying one demonstration episode and one task; query a single episode with WHERE episodeindex=0. - action0..5, observationstate0..5 (FIELD) - FLOAT 6-D SO-follower…
Apache TsFile version of axiboai/piper-umi-ee-pick-place. This is a pick-and-place demonstration dataset recorded on an AgileX Piper arm with a single 6-DoF end-effector (pipersingleeef6d) and packaged in the LeRobot v2.1 layout. Each row is one control step: a 10-dimensional end-effector state (position x, y, z, rotation as 6-D rot6d, and gripper) together with the 10-dimensional commanded action. The metadata records a single task (taskindex = 0). - Time (INT64, milliseconds) — round(timestamp 1000); restarts at 0 for each episode. - episodeindex (TAG) — episode / device dimension. - taskindex (TAG) — task dimension (constant 0 here). - frameindex (FIELD, INT64) — source frame counter…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Apache TsFile version of kjydb/cafe001. A teleoperation dataset recorded on a PS-A-03 500 arm and packaged in the LeRobot v2.1 layout. Each row is one control step: a 13-dimensional Cartesian state (position x, y, z, orientation rx, ry, rz, linear velocities x.vel, y.vel, z.vel, angular velocities pitch.vel, roll.vel, yaw.vel, and gripper.status) and the matching 13-dimensional commanded action. The metadata records a single task (taskindex = 0). - Time (INT64, milliseconds) — round(timestamp 1000); restarts at 0 for each episode. - episodeindex (TAG) — episode / device dimension. - taskindex (TAG) — task dimension (constant 0 here). - frameindex (FIELD, INT64) — source frame counter…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Episodes 0 to 9 of DreamMachines/actuatorunboxingfullboxcropDall (revision 6a8bded63372), with the street seen behind the robot cell replaced by a measuring lab. The robot, table, box, tray and actuators are the original pixels. Actions, joint states, timestamps, task and the intervention columns are copied unchanged. Only the three camera videos differ. - 10 episodes, 5,308 frames at 50 fps Same ten episodes with other backgrounds: - kiteml/dm-actuator-unboxing-bg-aircraft-workshop: an aircraft workshop - kiteml/dm-actuator-unboxing-bg-hangar: an aircraft hangar 1. SAM 3 finds the background (street, buildings, sky, cars, passers-by) and the foreground (robot, gripper, box, tray, actuator…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Apache TsFile version of nduque/evalrobustnesse93fullfklocal240k. This is an evaluation-robustness teleoperation dataset recorded on a Koch robot arm and packaged in the LeRobot v2.1 layout. Each row is one control step: a 6-dimensional joint/gripper state, the 6-dimensional commanded action, and a 4-dimensional local end-effector pose. The source dataset card does not name the manipulation task; the metadata records a single task (taskindex = 0). - Time (INT64, milliseconds) — round(timestamp 1000); restarts at 0 for each episode. - episodeindex (TAG) — episode / device dimension. - taskindex (TAG) — task dimension (constant 0 here). - frameindex (FIELD, INT64) — source frame counter…
Exact data, public pretrained initialization, and offline Linux x8664 environment for 12 independent RMBench tasks. This snapshot preserves the rmbenchfastwamgtcovercurrentgbs64200k training settings: seed 42, global batch 64, 200000 optimizer steps, current observations, ground-truth subtask text, 32 actions, and subtask-boundary terminal hold. 600 demonstrations / 416711 frames, with the exact original LeRobot episode order and 46 cached GT subtask prompts. Public Wan 2.2, VAE, T5 and ActionDiT initialization are included. Previously trained project checkpoints are excluded. Repository-relative paths match the accompanying code: https://github.com/exploreing2333/fastwam-rmbench-training…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Image-in, trajectory-out navigation scenarios with 3D ground truth, generated in simulation with Habitat-Sim in the houses of HSSD (Habitat Synthetic Scenes Dataset, Khanna et al. 2023). Same schema and evaluator as the real-robot conversions JRDB); the CODa card documents every column, the coordinate frame, the point and map formats and the evaluation protocol. This set is the large training counterpart of those small real-world test sets. Code: - Input, the past. Rendered forward-camera images (640 x 480, 90° horizontal field of view, 1.2 m above the floor) of the 10 past steps and the current step, the agent's own past positions, and the - Ground truth, the future. The shortest navmesh…
Kinex Benchmark Unified media and evidence storage for the Kinex Benchmark Space. The current index compares Codex + GPT-6 Astra / high and Kinex + GPT-6 Astra / high, with 84 selected seed-0 episodes per combination. Claude Code + Opus 5.5 / high and Kinex + Opus 5.5 / high are planned and have no results yet. Pending tasks remain pending; native verdicts, measured instructions, budgets and token counters are unchanged. Location data.json Current 164-task
is a large-scale, fully synthetic, ground-truth-complete benchmark for the two classical sensor-calibration problems of robotics, hand-eye calibration, A X = X B, and simultaneous robot-world / hand-eye calibration, A X = Y B, together with their spatio-temporal extension (unknown clock offset between the robot and the camera). Instead of images, every sequence consists of two time-stamped pose streams, the robot flange trajectory (from forward kinematics of real manipulator models) and the pose of the observed reference in the camera, both as exact ground truth and as measurements corrupted by physically motivated noise models. This isolates the numerical calibration problem from image…
