video video 11.4 66 |
|---|
RoboValue Training Data
English · 简体中文
Expert demonstrations for robotic value models, covering real-world and simulated manipulation tasks. Each domain has two demonstration-access settings: Full-Shot for training or fine-tuning, and One-Shot for a single task-specific reference or adaptation example.
The common release format is LeRobot Dataset v3.0: per-step Parquet data, camera videos, and task/episode metadata. Robot platforms, cameras, and state/action schemas are specified separately for each subset.
Available now: Real-World Full-Shot — 2,000 episodes across 20 tasks. The other three subsets have reserved directories and defined sizes, but do not yet contain trajectory files.
Real-world data · Simulation · Quick start · Repository files
1. Dataset Overview
| Domain | Tasks | Full-Shot | One-Shot |
|---|---|---|---|
| Real-World | 20 | 2,000 | 20† |
| Simulation | 15 | 1,500† | 15† |
Counts are episodes, not video files or individual camera views. Full-Shot provides 100 episodes per task; One-Shot selects one episode per task from the same Full-Shot pool, rather than adding independently collected demonstrations. † denotes the defined size of a reserved subset.
Which setting should I use?
- Full-Shot: learn or fine-tune task-conditioned value/progress representations using the full demonstration collection.
- One-Shot: condition a model on one reference demonstration, or perform task-specific adaptation under a one-example budget.
This repository contains standard in-domain training demonstrations, not the paper's diagnostic or OOD evaluation trajectories. It does not supply extra scalar reward or value-target labels; your training method must define its own supervision or target construction.
2. Real-World
2.1 Overview
Real robot demonstrations collected with G2A, covering 20 manipulation tasks. The tasks include object placement and sorting, household organization, drawers, repeated presses, kitchen activities, and packaging. Choose this domain when you need value-model training or reference demonstrations grounded in real visual observations and physical interaction.
Task catalog: all 20 tasks and short descriptions
Indices match the training data's task_index. Display names and descriptions follow Appendix D of the RoboValue paper; the original task strings in meta/tasks.parquet are retained.
| Index | Task and description |
|---|---|
| 0 | Transfer Blocks Using Plate Load red, green, then blue blocks onto a plate and pour them into the box. |
| 1 | Enter Keypad Number Sequence Enter 2-8-2-6-8, releasing and returning between presses. |
| 2 | Set the Table Take the plate, bowl, fork and knife from the crate and set the table. |
| 3 | Arrange Seasoning Jars Place three seasoning jars on the two-tier rack. |
| 4 | Load Plastic Plates Into Rack Insert three plastic plates into the rack. |
| 5 | Press Button Five Times Press the button five times, releasing and returning after each press. |
| 6 | Place Pot On Induction Cooker Place the pot on the cooker, store the spatula and cover the pot. |
| 7 | Organize Desktop Workstation Store the pen and eraser, close the drawer, switch off and unplug the lamp. |
| 8 | Organize Books And Pen Align one book; place the pen in the other, close it and stack it. |
| 9 | Arrange Family Tea Set Arrange the cup, teapot/lid, canister and tongs in their target areas. |
| 10 | Operate Drawers In Sequence Open the drawers bottom-to-top, then close them top-to-bottom. |
| 11 | Sort Blocks in Order Place red, yellow and blue blocks on their designated drawer units, in order. |
| 12 | Fill And Place Pen Holder Insert four pens into the holder and place it upright. |
| 13 | Organize Cosmetics Storage Box Store the lipstick and eyebrow pencil, close the drawer, then place the serum and cleansing oil. |
| 14 | Strike Drum Five Times Pick up the drumstick, strike five times and return it. |
| 15 | Apply Dish Soap Repeatedly Hold the cloth beneath the outlet and pump dish soap five times. |
| 16 | Season Meat Skewers Repeatedly Season the meat skewers with the shaker five times. |
| 17 | Pack Items For Shipping Pack the mouse, then the wired earphones, and cover both with bubble wrap. |
| 18 | Arrange Flowers and Reed Diffuser Arrange three flowers, place the diffuser bottle and insert the reed. |
| 19 | Prepare Breakfast Place the buns in their designated positions, add a steamer tier and cover it. |
2.2 Full-Shot
Use for training or fine-tuning with all 100 demonstrations per task. This is the currently available real-world training subset.
Download and read · Browse files · Feature definitions
Key statistics
| Item | Value |
|---|---|
| Coverage | 20 tasks × 100 episodes = 2,000 episodes |
| Time steps / sampling | 1,530,987 at 30 FPS |
| Trajectory duration | ≈14.18 hours total; ≈25.52 seconds per episode on average |
| Cameras | 3 RGB views: head, left wrist, right wrist |
| Resolution | Head 640 × 400; each wrist 640 × 528 |
| Storage | ≈23.23 GB; Parquet + HEVC video |
| Split | train: 0:2000 |
Duration and time steps count the trajectory once, not three times for the cameras.
What does one trajectory contain?
An episode is a complete task demonstration: task/episode identifiers, per-step robot observations and actions, and three camera streams. Episode metadata records its length, data-row boundaries, and video segments. In v3.0, episodes may share chunk files; a file is not necessarily one episode.
| Signal | Representation |
|---|---|
action |
52 float32 values: joint targets and end-effector target poses, plus gripper, head, waist, and base-control fields |
observation.state |
290 float32 values: joint and end-effector state, including position, orientation, velocity, torque/force, and auxiliary fields |
| RGB observations | Head, left-wrist, and right-wrist images; video features use CHW metadata |
| Identity and time | episode_index, task_index, frame_index, and per-step timestamp in seconds |
End-effector action targets are computed by forward kinematics from joint actions, not independent teleoperation pose commands. End-effector positions use meters in base_link; orientations use XYZW quaternions. Gripper values are not normalized. Select the representation your model expects; do not assume that all 52 action values are joint angles.
Action reference: all 52 dimensions
Python slices are stop-exclusive.
| Slice | Meaning |
|---|---|
[0:2] |
Left/right effector target opening or stroke |
[2:8] |
Arm end-effector target XYZ positions, m, base_link |
[8:16] |
Arm end-effector target XYZW orientations, base_link |
[16:30] |
Arm joint target positions, rad |
[30:33] |
Head joint target positions, rad |
[33:38] |
Waist joint target positions, rad |
[38:44] |
Base target XYZ linear and XYZ angular velocity |
[44:48] |
Left reference-pose XYZW quaternion |
[48:52] |
Right reference-pose XYZW quaternion |
The last two reference orientations are distinct from the end-effector target orientations. Fields with dimensions: 0 and indices: [] do not occupy the vector. Full definitions are in meta/info.json.
State reference: key fields in the 290-dimensional vector
| Slice | Meaning |
|---|---|
[0:2] |
Left/right effector state |
[2:26] |
End-effector force/torque fields |
[26:32] |
Arm end-effector XYZ positions, m, base_link |
[32:56] |
End-effector linear/angular velocity fields |
[56:64] |
Arm end-effector XYZW orientations, base_link |
[108:122] |
Arm joint positions, rad |
[122:136] |
Arm joint torques, Nm |
[136:150] |
Arm joint velocities, rad/s |
[164:176] |
Head joint positions, velocities, efforts, and modes |
[176:196] |
Waist joint positions, velocities, efforts, and modes |
[196:203] |
Robot position and orientation in odom |
This is a key-field summary, not the entire vector. Additional pose, mode, operator, camera-extrinsic, and reference fields are defined in meta/info.json.
2.3 One-Shot
Use for one-reference conditioning or task-specific adaptation. The setting selects one episode for each of the same 20 real-world tasks, for 20 episodes in total. These are selected from Full-Shot, not an additional task collection.
The One-Shot directory is reserved; its trajectory files and episode selection are not yet released. Until then, the Full-Shot quick start is the available entry point. Choosing your own single episode does not reproduce the official One-Shot selection.
3. Simulation
3.1 Overview
The simulation domain covers 15 manipulation tasks. Choose it for value-model experiments and demonstration conditioning in simulated environments. It follows the same Full-Shot/One-Shot organization and LeRobot Dataset v3.0 release specification, but its robot, cameras, and state/action schema must be read from its own subset metadata rather than inferred from G2A.
3.2 Full-Shot
15 tasks × 100 demonstrations = 1,500 episodes. Intended for training or fine-tuning with the complete simulation demonstration pool. The directory is reserved; trajectory statistics and the task/schema catalog will accompany the data release.
3.3 One-Shot
15 tasks × 1 demonstration = 15 episodes. Intended for single-reference conditioning or task-specific adaptation, using one selection per task from simulation Full-Shot. The directory is reserved. There is no simulation subset to load yet.
4. Repository Layout
README.md English dataset card
README_zh.md Chinese dataset card
subset_catalog.json Machine-readable subset index
data/
├── real-world/
│ ├── full-shot/ Available LeRobot dataset root
│ └── one-shot/ Reserved
└── simulation/
├── full-shot/ Reserved
└── one-shot/ Reserved
A released subset directory is the LeRobot root; the repository root is not. The reserved directories currently contain only a README.
Format reference: files inside the real-world Full-Shot subset
| File or directory | Purpose |
|---|---|
data/chunk-*/file-*.parquet |
Per-step state, action, identifiers, and timestamps |
videos/<camera>/chunk-*/file-*.mp4 |
RGB camera video streams |
meta/info.json |
v3.0 version, features, shapes, sampling rate, and path templates |
meta/tasks.parquet |
Original task strings and their task indices |
meta/episodes/chunk-*/file-*.parquet |
Episode length, task, data boundaries, video offsets, and per-episode statistics |
meta/stats.json |
Dataset-level feature statistics |
provenance/ |
Episode-source correspondence and data-conversion definitions |
checksums.sha256 |
File integrity manifest |
Camera feature keys are observation.images.cam_high, observation.images.cam_left_wrist, and observation.images.cam_right_wrist. Episode data boundaries are dataset_from_index and dataset_to_index (stop-exclusive); video timestamps are stored separately for each camera.
5. Download and Reading
Quick start — currently available Real-World Full-Shot. Public downloads require no token. The full subset is approximately 23.23 GB, including videos.
Use Python 3.10+ and a v3.0-compatible reader. The example is verified with LeRobot 0.4.2 and PyAV 15.1.0; Dataset v3.0 is the data format, not a package named LeRobot 3.0.
pip install "lerobot==0.4.2" "av==15.1.0"
from pathlib import Path
from huggingface_hub import HfApi, snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset
repo_id = "wzlllong/value-model-benchmark-training-data"
prefix = "data/real-world/full-shot"
revision = HfApi(token=False).dataset_info(repo_id).sha
snapshot = snapshot_download(
repo_id=repo_id,
repo_type="dataset",
revision=revision,
allow_patterns=[f"{prefix}/**"],
token=False,
)
dataset = LeRobotDataset(
repo_id=repo_id,
root=Path(snapshot) / prefix,
revision=revision,
video_backend="pyav",
)
sample = dataset[0]
print(dataset.num_episodes, dataset.num_frames)
print(sample["action"].shape, sample["observation.state"].shape)
Expected: 2,000 episodes, 1,530,987 time steps, action shape [52], and observation.state shape [290]. dataset[0] reads one time step, not an entire episode. LeRobot resolves the corresponding camera frames from video metadata.
The example pins one repository revision for both download and reading, and downloads only the chosen subset. Do not replace the prefix with a reserved directory: there are no trajectory files to load there yet.
License: no data license has been specified. Confirm applicable terms before use or redistribution.
- Downloads last month
- 242