wzlllong/value-model-benchmark-training-data

RoboValue Training and Small-Scale Evaluation Data

Training Demonstrations · Small-Scale Real-World & Simulation Public Test Sets

English · 简体中文

RoboValue provides expert demonstrations for robotic value models, covering 15 simulation tasks and 20 real-world tasks. Both domains offer Full-Shot and One-Shot subsets in the shared LeRobot Dataset v3.0 format.

All four training subsets are publicly available. Two small-scale public diagnostic test sets are organized separately under evaluation/, with 30 real-world trajectories and 30 simulation trajectories. Each domain has four tasks and 90 synchronized camera videos. Together the test subsets contain 60 trajectories, 44,647 time steps, and 180 videos. The paper and current leaderboard report Zero-Shot and One-Shot evaluations.

Overview · Simulation · Real-World Training · Public Test Sets · Metrics · Repository · Download & Read

📊 1. Dataset Overview

Domain	Platform	Tasks	Full-Shot (episodes)	One-Shot (episodes)
🟦 Simulation	ARX	15	1,500	15
🟧 Real-World	AgiBot Genie02	20	2,000	20

Setting	Demonstrations	Use
▦ Full-Shot	100 per task	Training or fine-tuning on the complete expert collection
◉ One-Shot	1 per task	Reference conditioning or task-specific adaptation

Across all four training subsets, the release contains 3,535 episodes, 2,348,375 synchronized time steps, and 10,605 camera videos. The two One-Shot subsets account for 35 episodes, 24,442 time steps, and 105 camera videos.

HF preview: The explicit real-world-test-30/test and simulation-test-30/test configurations each contain one row per test trajectory (30 rows). Select a configuration to browse its task, trajectory type, applicable metrics, and three camera video paths. Real-world is the default. These catalogs are browsing indices; native per-frame model input is read with LeRobot. Auxiliary JSON reports and navigation documents are excluded from the preview.

Public Test Sets

Training demonstrations and public test trajectories have separate directories and purposes:

Collection	Location	Scale	Purpose
Simulation / real-world training	training-data/<domain>/{full-shot,one-shot}	Four training subsets listed above	Expert demonstrations for training or reference conditioning
Real-world public test set	evaluation/real-world/test-30	4 tasks / 30 trajectories / 27,498 time steps / 90 videos	Small-scale In-Domain diagnostic evaluation
Simulation public test set	evaluation/simulation/test-30	4 tasks / 30 trajectories / 17,149 time steps / 90 videos	Small-scale In-Domain diagnostic evaluation

The real-world subset contains 12 desk, 12 flower-and-diffuser, 3 drum, and 3 button trajectories. The simulation subset contains 12 pen-holder, 12 mug-hanging, 3 press-by-number, and 3 block-swapping trajectories. Both supply images and English instructions, with evaluation annotations stored separately. Each subset has its own local task and episode indices, independent of training and of the other test subset. They provide partial diagnostic coverage rather than a complete leaderboard test release.

Coverage	Real-world	Simulation
TGA-CT, TGA-CF, VOC, Cycle-VOC, VS	Each uses 4 ED episodes	Each uses 4 ED episodes
SA / SIA	24 / 12 episodes	24 / 12 episodes
TRR	12 episodes / 4 three-role groups	12 episodes / 4 three-role groups
CSVC	12 episodes / 10 solution groups	12 episodes / 6 order groups
FPL	8 ER/EC episodes	Original FPL fields retained; no formal FPL cohort
VOC-MEM	6 complete counting histories	6 complete memory histories; no annotated counting-subtask intervals

See the real-world metric coverage and trajectory index and simulation metric coverage and trajectory index. Simulation ED/MS terminal labels and semantic alignment are user-attested; TRR/MEM terminal-label evidence retains the source status of pending independent verification. The simulation README documents these annotation evidence states.

Subtask annotations: All six training/test roots provide annotations/subtasks.json with S-numbered definitions, source references, and original label IDs where needed. Simulation One-Shot also provides 58 segments and 8,080 frame labels; both test roots provide all existing subtask segments. Other training roots provide reference definitions only. See Subtask reading for file formats, temporal-label availability, and loading examples.

Metric summary

These one-sentence definitions follow the paper’s §4.3 (pp. 9–12). ↑ Higher is better; FPL ↓ lower is better.

Metric	Brief definition
SA · Success Accuracy ↑	Measures the fraction of same-task successful/unsuccessful episode pairs in which the successful episode has the higher average terminal score.
TGA-CT · Task Grounding Accuracy - Cross-Task ↑	Measures how often the correct instruction gives a larger start-to-end value gain than every cross-task negative instruction for the same successful episode.
TGA-CF · Task Grounding Accuracy - Counterfactual ↑	Measures how often the correct instruction gives a larger start-to-end value gain than every counterfactual instruction changing the object, action, placement, or constraint.
SIA · Subtask Identification Accuracy ↑	Scores subtask recognition as the geometric mean of the independent judge’s probabilities for the correct subtask, based on the model’s free-form descriptions.
VOC · Value-Order Correlation ↑	Measures the Spearman rank correlation between predicted values and normalized time along successful trajectories.
Cycle-VOC ↑	Averages Spearman correlations with forward and reversed progress order when a successful trajectory is replayed forward then backward as one continuous sequence.
Memory-VOC (VOC-MEM) ↑	Applies VOC’s Spearman score to successful long trajectories with recurring visual states, using the history available to each model interface.
FPL · Failure-Point Localization ↓	Measures the absolute time error between the annotated error onset and the onset inferred from the largest detected value decline, normalized by trajectory duration.
TRR · Trajectory Recovery Reasoning ↑	Checks matched trajectories for the expected stage trends: failure down, continued error non-increasing, recovery attempt up, and successful/failed recovery outcome up/down.
VS · Value Stability ↑	Combines trend efficiency with non-flat time coverage to reward stable value updates while penalizing reversals and prolonged flat intervals.
CSVC · Cross-Solution Value Consistency ↑	Measures the consistency of local value gains for semantically matched subtasks across valid execution solutions.

The paper’s Memory-VOC is labeled VOC-MEM in this dataset, with membership key voc_mem. Stored labels and keys retain these names. VOC and VOC-MEM annotations are sampling anchors, not calibrated per-frame value scores.

🟦 2. Simulation Training

2.1 Overview

Simulation demonstrations use the ARX robot platform. They are based on the open-source RoboDojo dataset. Our team reviewed the data and recorded selected demonstrations again. They cover 15 manipulation tasks at 25 FPS.

Trajectory Contents · Shared by Full-Shot and One-Shot

In both Full-Shot and One-Shot, an episode is a complete task demonstration containing the task instruction, per-step states and actions, identifiers and timestamps, and three RGB camera streams. Episode metadata records its length, data-row range, and video segments for each camera.

Content	Representation
∿ Robot state	observation.state: float32[14]
➜ Robot action	action: float32[14]
▶ RGB observations	High, left-wrist, and right-wrist views, each 640 × 480
☰ Instruction and identity	Task instruction, task_index, episode_index, frame_index
◷ Time	timestamp in seconds; 25 FPS

Task List · 15 tasks

The simulation dataset covers the following 15 tasks.

Index	Task
0	Arrange Largest Number
1	Fill Pen Holder
2	Fold Clothes
3	Hang Mugs
4	Insert Tubes
5	Make Toast
6	Organize Table
7	Pack Objects Into Box
8	Play Stacking Toy
9	Press By Number
10	Put Bottles Into Dustbin
11	Stack Blocks
12	Store Laptop And Headphones
13	Swap Blocks
14	Sweep Blocks

Full-Shot task metadata · One-Shot task metadata

State Reference · 14 dimensions

observation.state · float32[14]

The first seven entries are left-side fields; the last seven are right-side fields.

Dimension indices (0-based)	Field names
0–6	left_joint_0 … left_joint_6
7–13	right_joint_0 … right_joint_6

Full-Shot metadata · One-Shot metadata

Action Reference · 14 dimensions

action · float32[14]

The first seven entries are left-side fields; the last seven are right-side fields.

Dimension indices (0-based)	Field names
0–6	left_joint_0 … left_joint_6
7–13	right_joint_0 … right_joint_6

Full-Shot metadata · One-Shot metadata

Camera Reference · 3 RGB streams

Loaded images use CHW (channels, height, width) order.

Camera key	Loaded shape
observation.images.cam_high	[3, 480, 640]
observation.images.cam_left_wrist	[3, 480, 640]
observation.images.cam_right_wrist	[3, 480, 640]

Full-Shot metadata · One-Shot metadata

▦ 2.2 Full-Shot

15 tasks × 100 demonstrations = 1,500 episodes.

▦ Time steps	▶ Videos	▣ Size	◷ Total duration	⑂ Training split
789,612	4,500 · AV1 / H.264	≈45.47 GB	≈8.77 hours	train: 0:1500

Browse files → · Download and read

◉ 2.3 One-Shot

15 tasks × 1 demonstration = 15 episodes. Uses the same task definitions as Full-Shot.

▦ Time steps	▶ Videos	▣ Size	◷ Total duration	⑂ Training split
8,080	45 · AV1	≈0.50 GB	≈5.39 minutes	train: 0:15

Browse files → · Task–episode–row mapping · Download and read

Subtask annotations: The separate adaptation supplement retains 58 subtask segments and 8,080 frame labels in annotations/subtasks.json, annotations/subtask_segments.parquet, and annotations/frame_labels.parquet. These files are separate from training model observations and test-set labels; reading instructions are linked above.

🟧 3. Real-World Training

3.1 Overview

Real-world demonstrations are collected with the AgiBot Genie02 robot. They cover 20 manipulation tasks, including object placement and sorting, household organization, drawers, repeated presses, kitchen activities, and packaging, at 30 FPS.

Trajectory Contents · Shared by Full-Shot and One-Shot

In both Full-Shot and One-Shot, an episode is a complete task demonstration containing the task instruction, per-step robot observations and actions, identifiers and timestamps, and RGB videos from the head, left wrist, and right wrist. Episode metadata records its length, data-row range, and video segments for each camera.

Content	Representation
∿ Robot state	observation.state: float32[290], including joint/end-effector state and auxiliary fields
➜ Robot action	action: float32[52], including joint targets, end-effector target poses, and other control fields
▶ RGB observations	Head 640 × 400; each wrist 640 × 528
☰ Instruction and identity	Task instruction, task_index, episode_index, frame_index
◷ Time	timestamp in seconds; 30 FPS

Task List · 20 tasks

The real-world dataset covers the following 20 tasks.

Index	Task
0	Transfer Blocks Using Plate
1	Enter Keypad Number Sequence
2	Set the Table
3	Arrange Seasoning Jars
4	Load Plastic Plates Into Rack
5	Press Button Five Times
6	Place Pot On Induction Cooker
7	Organize Desktop Workstation
8	Organize Books And Pen
9	Arrange Family Tea Set
10	Operate Drawers In Sequence
11	Sort Blocks in Order
12	Fill And Place Pen Holder
13	Organize Cosmetics Storage Box
14	Strike Drum Five Times
15	Apply Dish Soap Repeatedly
16	Season Meat Skewers Repeatedly
17	Pack Items For Shipping
18	Arrange Flowers and Reed Diffuser
19	Prepare Breakfast

Full-Shot task metadata · One-Shot task metadata

State Reference · 290 dimensions

observation.state · float32[290]

Dimension indices (0-based)	Meaning
0–1	Left/right effector state
2–25	End-effector force/torque fields
26–31	Arm end-effector XYZ positions, m, base_link
32–55	End-effector linear/angular velocity fields
56–63	Arm end-effector XYZW orientations, base_link
108–121	Arm joint positions, rad
122–135	Arm joint torques, Nm
136–149	Arm joint velocities, rad/s
164–175	Head joint positions, velocities, efforts, and modes
176–195	Waist joint positions, velocities, efforts, and modes
196–202	Robot position and orientation in odom

The table lists key fields; the full 290-dimensional definition is in the metadata below.

Full-Shot metadata · One-Shot metadata

Action Reference · 52 dimensions

action · float32[52]

End-effector action targets are computed by forward kinematics from joint actions. Positions use meters in base_link, and orientations use XYZW quaternions. Grippers use their original opening or stroke values. The 52-dimensional action vector includes end-effector poses, joint targets, and other control fields; dimension indices are listed below.

Dimension indices (0-based)	Meaning
0–1	Left/right effector target opening or stroke
2–7	Arm end-effector target XYZ positions, m, base_link
8–15	Arm end-effector target XYZW orientations, base_link
16–29	Arm joint target positions, rad
30–32	Head joint target positions, rad
33–37	Waist joint target positions, rad
38–43	Base target XYZ linear and XYZ angular velocity
44–47	Left reference-pose XYZW quaternion
48–51	Right reference-pose XYZW quaternion

The last two reference orientations are distinct from the end-effector target orientations. Full field definitions are in the metadata below.

Full-Shot metadata · One-Shot metadata

Camera Reference · 3 RGB streams

Loaded images use CHW (channels, height, width) order.

Camera key	Loaded shape
observation.images.cam_high	[3, 400, 640]
observation.images.cam_left_wrist	[3, 528, 640]
observation.images.cam_right_wrist	[3, 528, 640]

Full-Shot metadata · One-Shot metadata

▦ 3.2 Full-Shot

20 tasks × 100 demonstrations = 2,000 episodes.

▦ Time steps	▶ Videos	▣ Size	◷ Total duration	⑂ Training split
1,534,321	6,000 · HEVC	≈23.13 GB	≈14.21 hours	train: 0:2000

Browse files → · Download and read

◉ 3.3 One-Shot

20 tasks × 1 demonstration = 20 episodes. Uses the same task definitions as Full-Shot.

For 19 tasks, the selected trajectory has the smallest original episode_id among the task's 100 Full-Shot source trajectories. Arrange Family Tea Set uses 52375079, the smallest original ID in the 27-episode tea refresh batch. Original IDs and source conversion indices are recorded separately in the mapping below.

▦ Time steps	▶ Videos	▣ Size	◷ Total duration	⑂ Training split
16,362	60 · HEVC	≈0.22 GB	≈9.09 minutes	train: 0:20

Browse files → · Task–episode–row mapping · Download and read

🗂️ 4. Repository Layout

README.md                         English dataset card
README_zh.md                      Chinese dataset card
subset_catalog.json               Subset index
training-data/
├── simulation/
│   ├── full-shot/                1,500 episodes · 15 tasks
│   └── one-shot/                 15 episodes · 1 per task
└── real-world/
    ├── full-shot/                2,000 episodes · 20 tasks
    └── one-shot/                 20 episodes · 1 per task
evaluation/
├── simulation/test-30/           30 trajectories · 4 tasks
│   └── meta/, data/, videos/, annotations/, README files and catalog
└── real-world/test-30/            30 trajectories · 4 tasks
    ├── meta/                     Native LeRobot v3 metadata
    ├── data/                     Frame references
    ├── videos/                   Three synchronized camera streams
    ├── annotations/              Separate evaluation labels
    ├── episode_catalog.parquet   One browsing row per trajectory
    ├── README.md / README_zh.md   Task and metric documentation
    └── checksums.sha256           File integrity manifest

Each released training or public test subset directory is an independent LeRobot dataset root (root). Both test subsets have frame references, RGB videos, and English instructions, with evaluation labels alongside them; they have no robot state/action vectors. The primary view is the head camera for real-world and the front camera for simulation.

The repository-level training directory is training-data/. Each LeRobot root retains its native data/ frame-table directory; for example, Simulation One-Shot frame tables are stored under training-data/simulation/one-shot/data/.

Training subset files and metadata

File or directory	Purpose
data/chunk-*/file-*.parquet	Per-step states, actions, identifiers, and timestamps
videos/<camera>/chunk-*/file-*.mp4	RGB camera video streams
meta/info.json	Format version, features, dimensions, FPS, and path templates
meta/tasks.parquet	Task instructions and indices
meta/episodes/chunk-*/file-*.parquet	Episode length, task, data-row boundaries, video offsets, and episode statistics
meta/stats.json	Dataset-level feature statistics
provenance/	Episode sources and task correspondence
checksums.sha256	File integrity manifest

Data rows span [dataset_from_index, dataset_to_index); video segments are recorded separately for each camera. Time steps and duration are counted once per episode.

Official LeRobot v3.0 format guide · Machine-readable subset catalog

↓ 5. Download and Reading

5.1 Choose a subset

Select one prefix to download only that subset. Sizes use decimal GB.

🗂️ Subset	↓ prefix	▦ Episodes / time steps	▣ Size
🟦 Simulation Full-Shot	training-data/simulation/full-shot	1,500 / 789,612	≈45.47 GB
🟦 Simulation One-Shot	training-data/simulation/one-shot	15 / 8,080	≈0.50 GB
🟧 Real-World Full-Shot	training-data/real-world/full-shot	2,000 / 1,534,321	≈23.13 GB
🟧 Real-World One-Shot	training-data/real-world/one-shot	20 / 16,362	≈0.22 GB
Real-World Public Test	evaluation/real-world/test-30	30 / 27,498	≈0.38 GB
Simulation Public Test	evaluation/simulation/test-30	30 / 17,149	≈0.595 GB

5.2 Download and read

This example loads the smaller Simulation One-Shot subset and returns states, actions, and all three camera frames. To switch training subsets, change prefix to the corresponding training path in the table. For the public test set, use the separate image-and-instruction example below. Public data can be downloaded anonymously, and this example automatically reuses an existing local Hugging Face login. If anonymous downloads are rate-limited, run hf auth login and retry.

The examples were tested with Python 3.12, LeRobot 0.4.2, and PyAV 15.1.0:

python -m pip install "lerobot==0.4.2" "av==15.1.0"

import os
from pathlib import Path

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo_id = "wzlllong/value-model-benchmark-training-data"
prefix = "training-data/simulation/one-shot"
revision = "ec8c2f2ae736bae8f19ffd9e0ca5cd16dbbe236e"
local_dir = Path(os.environ.get("ROBOVALUE_LOCAL_DIR", "./robovalue-data"))

snapshot = snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    revision=revision,
    allow_patterns=[f"{prefix}/**"],
    local_dir=local_dir,
)
dataset = LeRobotDataset(
    repo_id=repo_id,
    root=Path(snapshot) / prefix,
    revision=revision,
    video_backend="pyav",
)
sample = dataset[0]
print("episodes / frames:", dataset.num_episodes, dataset.num_frames)
print("action / state:", tuple(sample["action"].shape),
      tuple(sample["observation.state"].shape))
for camera in dataset.meta.camera_keys:
    print(camera, tuple(sample[camera].shape))

Output:

episodes / frames: 15 8080
action / state: (14,) (14,)
observation.images.cam_high (3, 480, 640)
observation.images.cam_left_wrist (3, 480, 640)
observation.images.cam_right_wrist (3, 480, 640)

Files are saved under ./robovalue-data. Set the ROBOVALUE_LOCAL_DIR environment variable to an existing download directory to reuse local files. Download and reading use the same fixed data revision.

5.3 Use a training batch

All Full-Shot tasks have exactly 100 episodes in the train split: 15 simulation tasks (1,500 episodes) and 20 real-world tasks (2,000 episodes). Each One-Shot subset has exactly one episode per task.

After the example above, use the LeRobot dataset directly with a PyTorch DataLoader:

from torch.utils.data import DataLoader

loader = DataLoader(dataset, batch_size=2, shuffle=True, num_workers=0)
batch = next(iter(loader))
states = batch["observation.state"]
actions = batch["action"]
images = {camera: batch[camera] for camera in dataset.meta.camera_keys}
instructions = batch["task"]
print("batch action / state:", tuple(actions.shape), tuple(states.shape))

Simulation One-Shot prints batch action / state: (2, 14) (2, 14). For real-world data, the corresponding shapes are (2, 52) and (2, 290).

batch["task"] retains the source task text from meta/tasks.parquet; task_index matches the task lists above.

Each batch contains states, actions, task instructions, and three camera views. Configure the model's input and output dimensions for the selected domain. The training method constructs value or progress targets from these expert demonstrations.

Each time step is one training item. For temporal windows, configure LeRobot delta_timestamps.

5.4 Integrity checks

Each subset contains checksums.sha256. Run this command from the subset directory:

shasum -a 256 -c checksums.sha256

5.5 Download and read a public test subset

Choose evaluation/real-world/test-30 or evaluation/simulation/test-30 to download one public test subset. The example is pinned to 578593ef28ee7a2d861ebf38162c928291b0d4a9, which contains both test domains; downloading and loading use the same revision. Training examples use the separately pinned revision containing the training-data/ directory.

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo_id = "wzlllong/value-model-benchmark-training-data"
prefix = "evaluation/real-world/test-30"
revision = "578593ef28ee7a2d861ebf38162c928291b0d4a9"
snapshot = snapshot_download(
    repo_id=repo_id, repo_type="dataset", revision=revision,
    allow_patterns=[f"{prefix}/**"], local_dir="./robovalue-data",
)
root = Path(snapshot) / prefix
dataset = LeRobotDataset(
    repo_id=repo_id, root=root, revision=revision,
    video_backend="pyav", tolerance_s=1e-4,
)
sample = dataset[0]
value_input = {
    "images": {camera: sample[camera] for camera in dataset.meta.camera_keys},
    "instruction": sample["task"],
}
print(dataset.num_episodes, dataset.num_frames)  # real-world: 30 27498; simulation: 30 17149

The native metadata retains train: 0:30 as its LeRobot container split. Both public test subsets are intended for testing, and the HF browsing index uses the semantic test split. Load evaluation annotations explicitly from root / "annotations"; do not pass labels, success flags, trajectory types, metric memberships, or frame references to the model. History-based models can use delta_timestamps within each episode. See the real-world README or simulation README for annotation joins, metric cohorts, and integrity checks. The catalogs share 12 columns; video_primary refers to head/front respectively, and the native camera keys preserve those view meanings.

Publication authorization: All data in this dataset were authorized before public release.