RevoBench-Refactor: ten-task reference batch

This batch contains 10 accepted simulated demonstrations, 29,170 frames at 60 Hz, and 140 synchronized videos (486.17 seconds of robot time). Each task contributes one complete episode. The robot is the new RevoTron with textured, calibrated Revo3 MT hands and the approved palm-facing 180° wrist fisheye cameras. The previous fingertip optical calibration is retained.

The format is LeRobot v2.1, validated with LeRobot 0.3.3 / PyAV. These are selected scripted reference executions, not a statistically measured success rate or a large training benchmark. The only split is train.

Task previews

Overview-camera frames from the ten selected recordings, arranged in episode order. View full resolution · Source frames.

Episodes

Episode	Task	Frames	Seed	Randomized layout
0	hammer	1403	0	no
1	plug_charger	2457	1	yes
2	tray_carry	1676	1	yes
3	pour_can	3720	1	yes
4	cup_hang	3892	1	yes
5	rotate_knob	2230	1	yes
6	insert_pipette	2551	1	yes
7	insert_peg	3033	1	yes
8	kitchen	5424	1	yes
9	power_drill	2784	4	yes

English task instructions are in meta/tasks.jsonl. meta/batch_provenance.json maps every episode to its original recording. Episode, global-frame, task and tactile-target indices were assigned by the shared exporter; the original per-task datasets remain preserved separately.

Recorded fields

Use meta/info.json for exact feature names, shapes and joint ordering.

Fields	Shape per frame / interpretation
observation.state, action	[56], measured joint positions and absolute joint commands in radians; observations precede the corresponding command
observation.images.head, .wrist_left, .wrist_right, .overview	Four RGB videos; the two wrist cameras use 180° fisheye projection
observation.images.tactile_{side}_{finger}	Ten optical tactile RGB videos, five per hand
observation.tactile_depth.{side}_{finger}	[1,240,320] numeric float32 optical overlap depth in metres
observation.tactile_marker.{side}_{finger}	[100,3] float32 marker displacement in DIP-local coordinates, metres
observation.tactile_marker_valid.{side}_{finger}	[100] fixed boolean calibration validity mask
observation.tactile_pressure.{side}_{region}	MT regional response vectors, 451 channels per hand
observation.tactile_tip_wrench.{side}_fingers	[5,6] normal plus friction forces and moments about the DIP origin, in DIP-local coordinates; N and N m
observation.tactile_target_id	[1] dataset-wide target ID; resolve through meta/tactile_targets.json

Sides are left and right; fingers are thumb, index, middle, ring, pinky. Pressure has thumb regions of 31+57 channels, four finger regions of 21+52 each, and 71 palm channels. It is a nonnegative response after column-normalized spatial diffusion, not Pa or N, and can exceed 1. Do not clip it to [0,1]. The tactile backend observes one explicitly identified target at a time, not all scene contacts. Per-episode target_mapping.json maps original raw target slots to dataset-wide IDs.

Gains are configuration, not action dimensions. Every task uses hand stiffness/damping 80/8, arm stiffness [600,600,600,450,450,300,300] and arm damping 30. Recording/control run at 60 Hz; physics runs at 480 Hz.

Reading the data

Authenticate with an account permitted to read this private dataset (hf auth login). Use the pinned LeRobot 0.3.3 reader environment documented in the code repository. The snippet downloads a fixed format tag, then loads local files:

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

repo_id = "DavidLXu/RevoBench-Refactor"
root = snapshot_download(repo_id, repo_type="dataset", revision="v2.1")
dataset = LeRobotDataset(
    repo_id,
    root=root,
    download_videos=False,
    video_backend="pyav",
    delta_timestamps={"action": [i / 60 for i in range(16)]},
)
sample = dataset[0]
print(sample["observation.state"].shape)  # torch.Size([56])
print(sample["action"].shape)             # torch.Size([16, 56])

For an immutable experiment, replace v2.1 with this dataset's full Hub commit SHA. Preserve episode boundaries and use action_is_pad when training with action chunks. The official reader has been checked at the start, middle and end of all ten episodes, including repeated final-action padding.

Evidence and scope

meta/validation_summary.json contains combined-dataset numeric and official-reader validation. The exporter decoded every video to check all frame counts and timestamps. meta/episodes/episode_XXXXXX/ preserves the original raw manifest, physical report, calibration hashes and target mapping. meta/acceptance/ contains the selected task acceptance, randomized-layout evidence and individual physical, motion and visual reviews. meta/file_manifest.json records payload sizes and SHA-256 hashes, excluding itself.

These recordings share optical model physical_geometry_optical_intervals_v1_candidate. Numeric depth represents the connected target overlap along a calibrated ray at its gel-surface endpoint; it is not measured normal silicone indentation. The tray uses its closed authored source surface rather than the cooked PhysX SDF grid. Real-sensor fidelity has not been established. In the selected tray episode, thumb contact is on the shell outside the calibrated gel, so zero thumb optical response is valid.

pour_can uses the versioned granular surrogate pour_can_opened_granular_v2_cavity_frame; its containment predicate remains provisional and does not assert that every particle reaches the bowl. insert_pipette ends with the inserted tube still held. power_drill demonstrates pickup, non-contact board alignment and upright replacement/release; it does not simulate spindle rotation or drilling.

Code and reproduction: RevoBench-Refactor at c264625, recording instructions, sensor contract, physical criteria.

Robot, task assets, textures and calibration retain their respective upstream provenance; no new blanket asset license is asserted here. See third-party notices. This upload includes recorded data and evidence, not the full simulation asset bundle.