A human-in-the-loop, multi-view bimanual robot manipulation dataset for tabletop spill cleanup: removing distractor objects and wiping spilled liquid from the surface with a sponge. Delivered in a LeRobot-compatible structure with synchronized video, state/action trajectories, segmentation masks, object pose signals, phase annotations, and success/failure labels.

Why this dataset is different

Most public manipulation datasets come from one of two sources: real-robot teleoperation farms, which are slow and expensive, or pure simulation, which is cheap but often weak for transfer. This sample comes from a third path:

Human-in-the-loop VR capture. A human performs the task in an immersive virtual environment using a standard VR headset. Their motion provides task intent, manipulation timing, bimanual coordination, and correction behavior, while the system retargets the demonstration onto a virtual Franka Panda robot embodiment.
Procedurally expanded with added visual domain randomization. Seed demonstrations are expanded into physics-consistent variations with changing object poses, distractors, mild occlusions, lighting conditions, camera configurations, object materials, and environment appearance.
Packaged for direct inspection and training. The output is delivered in a LeRobot-compatible structure, with synchronized multi-view video, dual-arm state and action streams, segmentation masks, object pose labels, phase-level annotations, quality scores, and success/failure metadata.

The result is human-seeded, scaled, and labeled robot-manipulation data that is closer to what policy training needs, without requiring every trajectory to be collected on a physical robot.

This public release is intentionally compact. It is meant as an inspection sample: robotics teams can evaluate the format, modalities, visual variation, annotation schema, bimanual trajectory quality, failure semantics, segmentation streams, and object-pose metadata before discussing larger productized skill packs.

Dataset summary

Property	Value
Episodes	50
Total frames	67,461
Total duration	2,248.70 seconds, about 37.5 minutes
Task	Remove distractor objects and wipe the spilled liquid from the tabletop with a sponge
Robot embodiment	Bimanual Franka Emika Panda setup, two 7-DoF arms + parallel grippers
Camera views	7 synchronized RGB streams
RGB video	30 FPS, H.264, 1280 x 960
Segmentation masks	4 per-frame PNG instance-segmentation streams
Object pose labels	Per-frame object-state stream with positions, orientation, linear velocity, and dirty-fraction metadata
Object-state stream	Per-frame `observation.object_poses` list with sponge, cup, and liquid_spill
Rigid object 6DoF labels	Sponge and cup include 3D position plus quaternion orientation
Cleanup metric	`liquid_spill.dirty_fraction`, range 0.0 to 1.0
Recorded success	Terminal cleanup quality plus failure semantics such as operator aborts; see `next.success` and `annotations.json`
Robot state	18-dimensional
Action vector	18-dimensional
Trajectories	Synchronized dual-arm robot state + action streams per frame
Outcome mix	27 success episodes, 23 failure episodes
Failure reasons	20 cleanup-incomplete failures, 3 operator-abort failures
Frozen frames	8,181 total frozen frames across the sample
Phase-level annotations	approach, grasp, transport, place, retract, clean_surface, clean_swipe_pass, collision, handover, task_attempt
Episode-level metadata	success/failure outcome, failure reason, duration, frozen-frame count, quality scores, derived metrics
Visual variation	Object pose, distractors, mild occlusions, lighting, camera configuration, object material, and environment appearance variation
Format	LeRobot-compatible Parquet + MP4 videos + PNG mask frames
License	Apache 2.0

What's new in this version

This updated sample keeps the same 50 episode IDs but expands the available modalities and refreshes the per-frame metadata:

Adds a seventh RGB camera stream: observation.images.left_cam_45.
Adds four per-frame segmentation-mask streams:
- observation.masks.left_cam_45_seg
- observation.masks.top_cam_seg
- observation.masks.wrist_cam_l_seg
- observation.masks.wrist_cam_r_seg
Adds observation.object_poses, a per-frame object-state stream with object IDs, 3D positions, quaternion orientations, finite-difference linear velocities, and remaining dirty-fraction values.
Updates all episode parquet files to include the expanded schema.
Updates annotations.json and the LeRobot metadata files under meta/.

What is included

Each episode bundles synchronized robot, video, mask, pose, and annotation signals:

Robot state trajectories: the full 18D bimanual robot state stream over time, covering left and right Panda arm joints plus gripper finger joints.
Action trajectories: the 18D control/action signal at each frame for the same bimanual motor ordering.
Multi-view RGB video: seven synchronized camera streams: front, left, right, top, left wrist, right wrist, and left 45-degree camera.
Segmentation masks: per-frame PNG instance-segmentation masks for top, left 45-degree, left wrist, and right wrist views, plus label-map metadata in the parquet rows.
Object pose stream: per-frame object IDs, position, orientation, linear velocity, and remaining dirty-fraction values for tracked task objects.
Objective cleanup metric: dirty_fraction for the liquid_spill object, allowing cleanup quality to be audited from the terminal frame.
Success/failure semantics tied to cleanup quality and execution status: terminal dirty_fraction <= 0.01 means 1% or less of the original spill remains, while recorded success also accounts for failure labels such as operator aborts.
Per-frame indexing: timestamp, frame index, episode index, global index, task index, terminal state, and terminal success flag.
Episode-level metadata: task identity, success/failure outcome, failure reason, duration, frozen-frame count, quality scores, and derived execution metrics.
Phase-level annotations: frame-range segment boundaries for object approach, grasp, transport, placement, retraction, wiping passes, collision events, handovers, and failed task attempts.
Failure semantics: selected episodes include cleanup-incomplete, operator-abort, collision, wrong-object, and task-attempt failure signals in annotations and metrics.

Camera views

observation.images.front_cam
observation.images.left_cam
observation.images.left_cam_45
observation.images.right_cam
observation.images.top_cam
observation.images.wrist_cam_l
observation.images.wrist_cam_r

All RGB videos are 1280 x 960 H.264 streams at 30 FPS with no audio.

Segmentation mask fields

observation.masks.left_cam_45_seg
observation.masks.top_cam_seg
observation.masks.wrist_cam_l_seg
observation.masks.wrist_cam_r_seg

Segmentation masks are stored as 8-bit grayscale PNG label maps and referenced from parquet rows.

The label map is:

Label	Class
0	background
1	left_franka_panda
2	right_franka_panda
3	sponge
4	liquid_spill
5	cup
6	table_surface

Each mask field is represented as a parquet struct with:

instance_segmentation: string path to a PNG mask frame
label_map: map<string,string>

Object pose field

observation.object_poses

Each row stores a list of object records with:

object_id
position: float32[3], meters, right-handed world frame
orientation: float32[4], quaternion_xyzw, right-handed world frame
velocity_linear: float32[3], meters/second, finite difference
dirty_fraction: float32, remaining_dirty_fraction in [0.0, 1.0]

The stream contains three task objects per frame:

sponge
cup
liquid_spill

The sponge and cup are interactable rigid objects and include 6DoF pose labels: 3D position plus quaternion orientation. The liquid_spill is represented as a task-state object: it includes position and the remaining dirty fraction, while its orientation and linear velocity may be null.

The objective cleanup metric is:

liquid_spill.dirty_fraction

This value is a fraction in [0.0, 1.0], where:

1.0 means 100% of the original dirty area remains.
0.01 means 1% remains.
0.0 means the spill is fully cleaned.

The terminal cleanup-quality criterion is:

terminal_dirty_fraction <= 0.01

The recorded episode outcome is stored in next.success on the final frame and in annotations.json. It should be treated as authoritative because it also captures failure semantics such as operator_abort.

Core trajectory fields

observation.state
action
timestamp
frame_index
episode_index
index
task_index
next.done
next.success

Parquet schema

Each episode parquet contains 21 columns:

timestamp
frame_index
episode_index
index
task_index
next.done
next.success
observation.state
action
observation.object_poses
observation.images.front_cam
observation.images.left_cam
observation.images.top_cam
observation.images.right_cam
observation.images.wrist_cam_l
observation.images.wrist_cam_r
observation.images.left_cam_45
observation.masks.top_cam_seg
observation.masks.wrist_cam_l_seg
observation.masks.wrist_cam_r_seg
observation.masks.left_cam_45_seg

next.done is true on the final frame of each episode. next.success is true only on the final frame of successful episodes.

State and action motor order

left_panda_joint1
left_panda_joint2
left_panda_joint3
left_panda_joint4
left_panda_joint5
left_panda_joint6
left_panda_joint7
left_panda_finger_joint1
left_panda_finger_joint2
right_panda_joint1
right_panda_joint2
right_panda_joint3
right_panda_joint4
right_panda_joint5
right_panda_joint6
right_panda_joint7
right_panda_finger_joint1
right_panda_finger_joint2

Annotation fields

episode_id
success
task_success
failure_reason
duration_sec
frozen_frames
phase_annotations
scores
derived
raw_measurements
scorer_id

The phase_annotations field contains phase names, frame ranges, execution quality, task-alignment labels, hand labels, and optional notes such as grasped object, released object, collision object, or operator-abort context.

The scores, derived, and raw_measurements fields provide quality and diagnostic metrics such as path efficiency, grasp precision, placement accuracy, temporal efficiency, motion smoothness, corrective movement score, kinematic headroom, composite score, confidence, path ratio, discontinuity count, and low-frequency motion power.

Loading examples

Read one episode parquet

from pathlib import Path
import pyarrow.parquet as pq

root = Path("path/to/dataset")
episode = root / "data/chunk-000/episode_000000.parquet"

table = pq.read_table(episode)
df = table.to_pandas()

print(df.columns)
print(df[["timestamp", "frame_index", "next.done", "next.success"]].tail())

Load a referenced segmentation mask

from pathlib import Path
from PIL import Image
import numpy as np
import pyarrow.parquet as pq

root = Path("path/to/dataset")
df = pq.read_table(root / "data/chunk-000/episode_000000.parquet").to_pandas()

mask_ref = df["observation.masks.top_cam_seg"].iloc[0]
mask = np.array(Image.open(root / mask_ref["instance_segmentation"]))

print(mask.shape, mask.dtype, sorted(np.unique(mask).tolist()))
print(mask_ref["label_map"])

Read the terminal cleanup metric

from pathlib import Path
import pyarrow.parquet as pq

root = Path("path/to/dataset")
df = pq.read_table(root / "data/chunk-000/episode_000001.parquet").to_pandas()

terminal_objects = df["observation.object_poses"].iloc[-1]
spill = next(obj for obj in terminal_objects if obj["object_id"] == "liquid_spill")

dirty_fraction = float(spill["dirty_fraction"])
cleanup_quality_pass = dirty_fraction <= 0.01
recorded_success = bool(df["next.success"].iloc[-1])

print("dirty_fraction:", dirty_fraction)
print("remaining_percent:", dirty_fraction * 100.0)
print("cleanup_quality_pass:", cleanup_quality_pass)
print("recorded_success:", recorded_success)

Repository structure

README.md
annotations.json
meta/
  info.json
  tasks.jsonl
  episodes.jsonl
  episodes_stats.jsonl
data/
  chunk-000/
    episode_000000.parquet
    episode_000001.parquet
    ...
videos/
  chunk-000/
    observation.images.front_cam/
      episode_000000.mp4
      episode_000001.mp4
      ...
    observation.images.left_cam/
      episode_000000.mp4
      episode_000001.mp4
      ...
    observation.images.left_cam_45/
      episode_000000.mp4
      episode_000001.mp4
      ...
    observation.images.right_cam/
      ...
    observation.images.top_cam/
      ...
    observation.images.wrist_cam_l/
      ...
    observation.images.wrist_cam_r/
      ...
    observation.masks.left_cam_45_seg/
      episode_000000/
        left_cam_45_seg.FinalImageStencil.0000.png
        left_cam_45_seg.FinalImageStencil.0001.png
        ...
    observation.masks.top_cam_seg/
      episode_000000/
        top_cam_seg.FinalImageStencil.0000.png
        top_cam_seg.FinalImageStencil.0001.png
        ...
    observation.masks.wrist_cam_l_seg/
      ...
    observation.masks.wrist_cam_r_seg/
      ...

Intended use

This sample is suitable for:

Inspecting the Exylos data format and annotation schema.
Testing LeRobot-compatible training and data-loading pipelines.
Quick imitation-learning experiments on a narrow bimanual spill-cleanup task.
Evaluating synchronized multi-view RGB, dual-arm state/action trajectories, segmentation masks, object poses, and phase-level annotations.
Inspecting visual domain randomization and procedural variation in a compact manipulation sample.
Reviewing success, cleanup-incomplete, operator-abort, collision, wrong-object, and task-attempt failure examples.
Prototyping perception-conditioned policies that use RGB, masks, object-state signals, or combinations of these modalities.

For larger production-scale skill packs, including broader object families, configurable embodiments, denser labels, custom evaluation logic, or higher episode volumes, visit exylos.ai or contact us directly.

Out-of-scope

This sample does not target a specific real-world deployment cell or production line.
This 50-episode sample does not include depth maps or RGB-D streams.
It does not include a held-out benchmark split tuned for leaderboard-style evaluation.
It is a compact inspection sample rather than a broad benchmark for all spill-cleanup behaviors.
The segmentation masks and object pose labels are provided for inspection, loading, and prototype training; downstream users should validate that the semantics match their own evaluation or policy-training requirements.

Validation notes

The local sample was checked for:

50 parquet files with matching episode IDs from episode_000000 through episode_000049.
350 RGB MP4 files, 7 camera streams per episode.
269,844 segmentation PNG files, 4 mask streams per frame.
All parquet references to segmentation mask assets resolve to existing files.
Sampled segmentation masks load as 8-bit grayscale images with labels 0..6.
observation.object_poses contains the expected task objects: sponge, cup, and liquid_spill.
Terminal cleanup quality, dirty_fraction <= 0.01, aligns with recorded success except for operator-abort semantics; next.success and annotations.json remain the authoritative outcome fields.

About Exylos

Exylos is an early-stage robotics data company. We capture human manipulation demonstrations in consumer VR and procedurally expand them into physics-consistent, transfer-oriented training episodes with visual domain randomization. Datasets are delivered in LeRobot-compatible structure or adapted to client pipelines.

If you are a robotics or applied-ML team and want to discuss a custom skill pack for your embodiment and task, reach out at contact@exylos.ai or visit exylos.ai.

Citation

If you use this dataset in research or in a public technical report, please cite it as:

@misc{exylos_bimanual_table_spill_cleanup_sample_2026,
  title        = {Exylos Bimanual Table Spill Cleanup Sample: A Multi-View, VR-Captured Sponge-Wiping Dataset},
  author       = {Exylos},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/ExylosAi/table_spill_cleanup_bimanual}},
  note         = {LeRobot-compatible dataset with RGB video, masks, object poses, and phase annotations}
}

License

Released under the Apache License 2.0. This sample is intentionally permissive so robotics and ML teams can inspect, load, test, and commercially evaluate the format without licensing friction. You are free to use this dataset for both research and commercial purposes, subject to the standard Apache 2.0 attribution requirements.

Contact

Website: exylos.ai
Email: contact@exylos.ai
LinkedIn: Exylos on LinkedIn

For questions specific to this dataset, including format, schema, or fields, please open a discussion in the Community tab on this repository.

Downloads last month: 19,804

Total file size:

22.5 GB