The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
_id: struct<$oid: string>
child 0, $oid: string
name: string
slug: string
version: string
created_at: struct<$date: string>
child 0, $date: string
last_modified_at: struct<$date: string>
child 0, $date: string
last_deletion_at: null
last_loaded_at: struct<$date: string>
child 0, $date: string
sample_collection_name: string
persistent: bool
media_type: string
group_media_types: struct<>
tags: list<item: null>
child 0, item: null
info: struct<>
app_config: struct<dynamic_groups_target_frame_rate: int64, grid_media_field: string, media_fallback: bool, medi (... 75 chars omitted)
child 0, dynamic_groups_target_frame_rate: int64
child 1, grid_media_field: string
child 2, media_fallback: bool
child 3, media_fields: list<item: string>
child 0, item: string
child 4, modal_media_field: string
child 5, plugins: struct<>
classes: struct<>
default_classes: list<item: null>
child 0, item: null
mask_targets: struct<>
default_mask_targets: struct<>
skeletons: struct<>
camera_intrinsics: struct<>
static_transforms: struct<>
sample_fields: list<item: struct<name: string, ftype: string, embedded_doc_type: string, subfield: string, fields: (... 318 chars omitted)
child 0, item: struct<name: string, ftype: string, embedded_doc_type: string, subfield: string, fields: list<item: (... 306 chars omitted)
child 0, name: string
child 1, ftype: string
child 2, embedded_doc_type: string
child 3, subfield: string
child 4, fields: list<item:
...
92 chars omitted)
child 0, qwen3vl_umap: struct<_id: struct<$oid: string>, _dataset_id: struct<$oid: string>, key: string, version: string, t (... 370 chars omitted)
child 0, _id: struct<$oid: string>
child 0, $oid: string
child 1, _dataset_id: struct<$oid: string>
child 0, $oid: string
child 2, key: string
child 3, version: string
child 4, timestamp: struct<$date: string>
child 0, $date: string
child 5, config: struct<cls: string, type: string, method: string, embeddings_field: string, points_field: null, simi (... 173 chars omitted)
child 0, cls: string
child 1, type: string
child 2, method: string
child 3, embeddings_field: string
child 4, points_field: null
child 5, similarity_index: null
child 6, model: null
child 7, model_kwargs: null
child 8, patches_field: null
child 9, num_dims: int64
child 10, num_neighbors: int64
child 11, metric: string
child 12, min_dist: double
child 13, seed: null
child 14, verbose: bool
child 6, results: struct<$oid: string>
child 0, $oid: string
child 7, results_meta: struct<>
evaluations: struct<>
runs: struct<>
active_label_schemas: list<item: null>
child 0, item: null
label_schemas: struct<>
frame_label_schemas: struct<>
reducer_blob: string
sample_ids: string
cls: string
label_ids: null
points: string
to
{'cls': Value('string'), 'reducer_blob': Value('string'), 'points': Value('string'), 'sample_ids': Value('string'), 'label_ids': Value('null')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
_id: struct<$oid: string>
child 0, $oid: string
name: string
slug: string
version: string
created_at: struct<$date: string>
child 0, $date: string
last_modified_at: struct<$date: string>
child 0, $date: string
last_deletion_at: null
last_loaded_at: struct<$date: string>
child 0, $date: string
sample_collection_name: string
persistent: bool
media_type: string
group_media_types: struct<>
tags: list<item: null>
child 0, item: null
info: struct<>
app_config: struct<dynamic_groups_target_frame_rate: int64, grid_media_field: string, media_fallback: bool, medi (... 75 chars omitted)
child 0, dynamic_groups_target_frame_rate: int64
child 1, grid_media_field: string
child 2, media_fallback: bool
child 3, media_fields: list<item: string>
child 0, item: string
child 4, modal_media_field: string
child 5, plugins: struct<>
classes: struct<>
default_classes: list<item: null>
child 0, item: null
mask_targets: struct<>
default_mask_targets: struct<>
skeletons: struct<>
camera_intrinsics: struct<>
static_transforms: struct<>
sample_fields: list<item: struct<name: string, ftype: string, embedded_doc_type: string, subfield: string, fields: (... 318 chars omitted)
child 0, item: struct<name: string, ftype: string, embedded_doc_type: string, subfield: string, fields: list<item: (... 306 chars omitted)
child 0, name: string
child 1, ftype: string
child 2, embedded_doc_type: string
child 3, subfield: string
child 4, fields: list<item:
...
92 chars omitted)
child 0, qwen3vl_umap: struct<_id: struct<$oid: string>, _dataset_id: struct<$oid: string>, key: string, version: string, t (... 370 chars omitted)
child 0, _id: struct<$oid: string>
child 0, $oid: string
child 1, _dataset_id: struct<$oid: string>
child 0, $oid: string
child 2, key: string
child 3, version: string
child 4, timestamp: struct<$date: string>
child 0, $date: string
child 5, config: struct<cls: string, type: string, method: string, embeddings_field: string, points_field: null, simi (... 173 chars omitted)
child 0, cls: string
child 1, type: string
child 2, method: string
child 3, embeddings_field: string
child 4, points_field: null
child 5, similarity_index: null
child 6, model: null
child 7, model_kwargs: null
child 8, patches_field: null
child 9, num_dims: int64
child 10, num_neighbors: int64
child 11, metric: string
child 12, min_dist: double
child 13, seed: null
child 14, verbose: bool
child 6, results: struct<$oid: string>
child 0, $oid: string
child 7, results_meta: struct<>
evaluations: struct<>
runs: struct<>
active_label_schemas: list<item: null>
child 0, item: null
label_schemas: struct<>
frame_label_schemas: struct<>
reducer_blob: string
sample_ids: string
cls: string
label_ids: null
points: string
to
{'cls': Value('string'), 'reducer_blob': Value('string'), 'points': Value('string'), 'sample_ids': Value('string'), 'label_ids': Value('null')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Dataset Card for WireFishing-M
WireFishing-M is a multimodal robotics dataset for deformable cable insertion, capturing a Franka Emika Panda 7-DOF arm equipped with an Allegro Robot Hand and DIGIT GelSight tactile sensor inserting seven types of cables into a transparent L-shaped PVC pipe. Each episode contains synchronized streams from four RGB cameras (bottom view, front view, pipe-side view, and tactile sensor) plus per-frame robot state (end-effector force/torque, pose, joint angles, and insertion success label). This FiftyOne dataset contains 14 MCAP episodes from the Mendeley representative subset, structured as multimodal samples with five synchronized channels per episode. The full dataset (≈3.9 million frames) is available on Harvard Dataverse.
Installation
pip install -U fiftyone
Usage
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub
dataset = load_from_hub("harpreetsahota/wirefishing-m")
session = fo.launch_app(dataset)
Dataset Details
Dataset Sources
- Repository: Harvard Dataverse — WireFishing-M-1 · WireFishing-M-2 · Mendeley Data subset
- Paper: Zhou et al., "WireFishing-M: A multimodal dataset for deformable cable insertion using tactile, visual, and proprioceptive sensing," Data in Brief, vol. 63, Dec. 2025, 112136. https://doi.org/10.1016/j.dib.2025.112136
- Demo: [More Information Needed]
Uses
Direct Use
WireFishing-M supports research in:
- Multimodal perception and sensor fusion — the four synchronized camera streams (bottom, front, pipe-side, tactile) and robot state provide rich input for early and late fusion architectures.
- Insertion success detection — automated binary success labels derived from pipe-side camera pixel counts enable training and evaluation of task-completion classifiers.
- Force estimation and contact modeling — end-effector force/torque time-series paired with tactile images support data-driven contact estimation.
- Manipulation policy learning — human demonstration trials with higher success rates and randomized trials with diverse failure modes cover both imitation learning and reinforcement learning scenarios.
- Generalization across cable types — seven cables varying in stiffness, diameter, surface texture, and weight support transfer learning and domain adaptation studies.
Out-of-Scope Use
This dataset was collected in a fixed laboratory environment using a single pipe geometry. It is not suitable as-is for benchmarking manipulation in unstructured environments or with pipe geometries other than the 1″ Sch 40 L-shaped PVC assembly used during collection.
Dataset Structure
Overview
This FiftyOne dataset contains 14 MCAP episodes (samples), one per trial in the Mendeley representative subset. Each sample's media_type is multimodal. Episodes are authored at 5 fps from the raw per-frame files; timestamps are synthesized as frame_index × 200 ms because the original ROS timestamps are not included in the Mendeley export.
The full WireFishing-M dataset on Harvard Dataverse contains approximately 3.9 million randomized-trial frames and 44,288 human-demonstration frames across multiple date-stamped sessions per cable type. The Mendeley subset provides one trial per cable per condition (14 total), with 675–1,000 frames per episode.
MCAP Channel Structure
Each MCAP episode contains nine synchronized channels (all protobuf):
| Topic | Schema | Tile | Resolution | Notes |
|---|---|---|---|---|
/camera/global/image_raw |
foxglove.CompressedImage |
Image | 1280 × 720 | Bottom view (Azure Kinect). Folder on disk: globel_view/ (dataset typo) |
/camera/local/image_raw |
foxglove.CompressedImage |
Image | 1920 × 1080 | Front view (Femto Mega). Paper states 1280 × 720; actual files are 1920 × 1080 |
/camera/inner/image_raw |
foxglove.CompressedImage |
Image | 640 × 480 | Pipe-side view (USB endoscope) |
/camera/tactile/image_raw |
foxglove.CompressedImage |
Image | 240 × 960 | Three DIGIT fingertips stacked vertically (index top, thumb middle, middle finger bottom). Paper states grayscale; files are RGB JPEG |
/robot/ee_force |
foxglove.Vector3 |
Plot | — | End-effector force (x=Fx, y=Fy, z=Fz) in N |
/robot/ee_torque |
foxglove.Vector3 |
Plot | — | End-effector torque (x=Tx, y=Ty, z=Tz) in N·m |
/robot/ee_position |
foxglove.Vector3 |
Plot | — | End-effector position (x, y, z) in m |
/robot/ee_orientation |
foxglove.Vector3 |
Plot | — | End-effector orientation (x=roll, y=pitch, z=yaw) in rad. Note: paper describes quaternion; empirical values indicate Euler RPY |
/robot/joint_states |
foxglove.JointStates |
Plot | — | 7 Panda joint angles (joint_1…joint_7, position field in rad) plus success joint (position = 1.0 for successful insertion, 0.0 for failure) |
Sample-Level Fields
| Field | FiftyOne type | Description |
|---|---|---|
filepath |
StringField |
Absolute path to the .mcap episode file |
cable_id |
IntField |
Cable number (1–7), inferred from folder name |
condition |
StringField |
Insertion condition: "human" (manually guided, higher success rate) or "random" (randomized start pose, diverse failure modes) |
trial |
IntField |
Trial number within the cable/condition folder (always 1 in this Mendeley subset) |
has_robot_data |
BooleanField |
True if the episode contains robot state channels. All 14 episodes in this subset have robot data |
size_mb |
FloatField |
MCAP file size in megabytes (342–525 MB per episode) |
n_frames |
IntField |
Number of synchronized frames in the episode |
n_success_frames |
IntField |
Number of frames where insertion_success == 1 |
success_rate |
FloatField |
Fraction of frames labeled as successful insertion |
max_force_magnitude |
FloatField |
Maximum end-effector force magnitude (N) across all frames: max(√(Fx²+Fy²+Fz²)) |
mean_force_magnitude |
FloatField |
Mean end-effector force magnitude (N) across all frames |
max_torque_magnitude |
FloatField |
Maximum end-effector torque magnitude (N·m) across all frames: max(√(Tx²+Ty²+Tz²)) |
ee_pos_range_x |
FloatField |
Range of end-effector x-position across the episode (m) |
ee_pos_range_y |
FloatField |
Range of end-effector y-position across the episode (m) |
ee_pos_range_z |
FloatField |
Range of end-effector z-position across the episode (m) |
clip_embedding |
VectorField |
512-dim CLIP ViT-B/32 embedding of the middle-frame global camera image, L2-normalized. Computed from openai/clip-vit-base-patch32 via Hugging Face Transformers |
qwen3vl_embedding |
VectorField |
6144-dim video embedding: Qwen3-VL-Embedding-8B embed_frames applied independently to 200 sampled frames (stride 5) from each of three camera channels (global, local, inner), then concatenated and L2-normalized. Captures temporal dynamics across three viewpoints simultaneously |
Brain Runs
| Brain key | Method | Description |
|---|---|---|
clip_umap |
UMAP (2D) | Dimensionality reduction of clip_embedding for visual exploration of episode similarity in the FiftyOne App Embeddings panel |
qwen3vl_umap |
UMAP (2D) | Dimensionality reduction of qwen3vl_embedding — multiview video-aware episode similarity |
Robot State Fields (source .npy mapping)
| MCAP channel | Field path | Source indices | Units | Notes |
|---|---|---|---|---|
/robot/ee_force |
.x, .y, .z |
[0–2] | N | End-effector force Fx, Fy, Fz |
/robot/ee_torque |
.x, .y, .z |
[3–5] | N·m | End-effector torque Tx, Ty, Tz |
/robot/ee_position |
.x, .y, .z |
[6–8] | m | End-effector Cartesian position |
/robot/ee_orientation |
.x, .y, .z |
[9–11] | rad | End-effector orientation (roll, pitch, yaw). Paper describes quaternion; empirical values indicate Euler RPY |
/robot/joint_states |
joints[0–6].position |
[12–18] | rad | Franka Emika Panda joint angles (7-DOF) |
/robot/joint_states |
joints[7].position (name=success) |
[19] | — | Binary insertion label: 1.0 = success, 0.0 = failure |
To plot force x in the FiftyOne viewer: add series /robot/ee_force.x. For joint angles: /robot/joint_states.joints[0].position through joints[6].position.
Indices [20–33] of the raw .npy arrays are not logged (zero for all episodes in this subset; not described in the paper).
Cable Types
cable_id |
Description | Full-dataset random frames | Full-dataset human frames |
|---|---|---|---|
| 1 | Flexible, Thin, Lightweight, Braided nylon cover | ≈2,900,000 | 5,962 |
| 2 | Rigid, Thick, Heavy-duty, No cover | 84,621 | 5,994 |
| 3 | Rigid, Thin, Lightweight, No cover | 57,285 | 6,668 |
| 4 | Less Flexible, Thin, Lightweight, Braided nylon cover | 169,602 | 6,170 |
| 5 | Less Rigid, Thin, Lightweight, No cover | 284,262 | 7,624 |
| 6 | Rigid, Thin, Lightweight, Braided nylon cover | 27,232 | 4,949 |
| 7 | Flexible, Thin, Lightweight, No cover | 384,475 | 7,141 |
Mendeley Subset Frame Counts (this FiftyOne dataset)
cable_id |
condition |
Frames | MCAP size |
|---|---|---|---|
| 1 | human | 1,000 | 471 MB |
| 1 | random | 1,000 | 474 MB |
| 2 | human | 914 | 450 MB |
| 2 | random | 1,000 | 521 MB |
| 3 | human | 675 | 342 MB |
| 3 | random | 1,000 | 506 MB |
| 4 | human | 1,000 | 493 MB |
| 4 | random | 1,000 | 510 MB |
| 5 | human | 1,000 | 480 MB |
| 5 | random | 1,000 | 524 MB |
| 6 | human | 755 | 369 MB |
| 6 | random | 1,000 | 524 MB |
| 7 | human | 839 | 408 MB |
| 7 | random | 1,000 | 525 MB |
Splits
No train/val/test split is provided. Filter by condition ("human" / "random") or cable_id (1–7) using FiftyOne views.
Dataset Creation
Curation Rationale
WireFishing-M was created to address the lack of multimodal benchmarks for deformable object manipulation in contact-rich environments. The wire insertion task is representative of real-world scenarios in construction, electrical assembly, and robotics. Collecting synchronized tactile, visual, proprioceptive, and force data across seven cable types and two collection conditions (human demonstration and randomized robot insertions) provides a resource covering both expert-like behavior and diverse failure modes.
Source Data
Data Collection and Processing
Data was collected at the ICIC Lab, University of Florida (1949 Stadium Rd, Weil Hall 360, Gainesville, FL 32611). A Franka Emika Panda 7-DOF manipulator on a Vention workstation was equipped with an Allegro Robot Hand (16-DOF) and a DIGIT GelSight tactile sensor on the index, middle, and thumb fingertips. The insertion target was a transparent L-shaped PVC pipe (1″ Sch 40 NSF-61).
Sensors and recording rates:
- Front camera: Femto Mega RGB-D, 1280 × 720 px, 5 fps
- Bottom camera: Microsoft Azure Kinect DK, 1280 × 720 px, 5 fps
- Pipe-side camera: 16.4 ft USB endoscope, 640 × 480 px, 5 fps
- Tactile sensor: DIGIT GelSight on three fingertips, concatenated to 240 × 960 px, ≈5 fps
- Robot state: Franka ROS interface (
/franka_state_controller/franka_states), 5 fps
All streams were timestamped via ROS Melodic on Ubuntu 18.04 and synchronized offline using software timestamps. No hardware synchronization was applied. Camera intrinsic and extrinsic parameters were calibrated with a checkerboard pattern; calibration files are not included in the dataset. Raw data was post-processed into per-frame .jpg and .npy files.
For randomized trials, the robot moved to a randomized initial pose and then applied incremental forward steps (0–2 cm along x-axis) with simultaneous small adjustments in y/z (±1 cm) and random pitch/yaw (±15°), resetting when a predefined insertion depth was exceeded. For human demonstration trials, a human physically guided the robot end-effector through the insertion.
Who are the source data producers?
Data was recorded by the authors at the University of Florida ICIC Lab. No third-party human subjects or social media data were involved.
Annotations
Annotation process
Insertion success was labeled automatically. The pipe-side camera image was converted to grayscale; if the pixel count below intensity 50 in a predefined region near the pipe outlet exceeded 40 pixels, the trial frame was labeled as a successful insertion (label = 1), otherwise as failure (label = 0). This labeling is stored as index [19] in the .npy robot state array and exposed as insertion_success in the /robot/state MCAP channel.
Who are the annotators?
Automated algorithm applied by the dataset authors. No human annotators.
Personal and Sensitive Information
The dataset contains no human subjects, biometric data, or personally identifiable information.
Citation
BibTeX:
@article{zhou2025wirefishing,
title = {WireFishing-M: A multimodal dataset for deformable cable insertion using tactile, visual, and proprioceptive sensing},
author = {Zhou, Tianyu and You, Hengxu and Xu, Fang and Du, Jing},
journal = {Data in Brief},
volume = {63},
pages = {112136},
year = {2025},
doi = {10.1016/j.dib.2025.112136}
}
APA:
Zhou, T., You, H., Xu, F., & Du, J. (2025). WireFishing-M: A multimodal dataset for deformable cable insertion using tactile, visual, and proprioceptive sensing. Data in Brief, 63, 112136. https://doi.org/10.1016/j.dib.2025.112136
More Information
- The Mendeley subset (this FiftyOne dataset) contains 14 episodes. The full dataset is on Harvard Dataverse: WireFishing-M-1 and WireFishing-M-2.
- Raw per-frame files in the Mendeley download use non-sequential frame indices. The
_Nsuffix across all five subfolders within a trial is the synchronization key: all files named*_N.*for the sameNare co-temporal. - The bottom-camera folder is named
globel_view(typo) in the downloaded dataset, notglobal_viewas described in the paper. - The raw
.npyrobot state arrays have shape(34,). Indices [20–33] are zero in the Mendeley subset and are not documented in the paper.
Dataset Card Authors
Harpreet Sahota
Dataset Card Contact
[More Information Needed]
- Downloads last month
- 9