Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
arm_pose: list<item: double>
child 0, item: double
depth_intrinsics: struct<d: list<item: double>, distortion_model: string, frame_id: string, height: int64, k: list<ite (... 48 chars omitted)
child 0, d: list<item: double>
child 0, item: double
child 1, distortion_model: string
child 2, frame_id: string
child 3, height: int64
child 4, k: list<item: double>
child 0, item: double
child 5, p: list<item: double>
child 0, item: double
child 6, width: int64
depth_stamp_ns: int64
depth_units: string
detections: struct<camera_frame: string, detections: list<item: struct<axis_direction_valid: bool, axis_elongati (... 1546 chars omitted)
child 0, camera_frame: string
child 1, detections: list<item: struct<axis_direction_valid: bool, axis_elongation: double, axis_end_width_ratio: double, (... 1316 chars omitted)
child 0, item: struct<axis_direction_valid: bool, axis_elongation: double, axis_end_width_ratio: double, axis_endpo (... 1304 chars omitted)
child 0, axis_direction_valid: bool
child 1, axis_elongation: double
child 2, axis_end_width_ratio: double
child 3, axis_endpoints_base_m: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
child 4, axis_length_m: double
child 5, axis_mask_pixels: int64
child 6, axis_negative_end_width_px: double
child 7, axis_positive_end_width_px: double
c
...
d 7, match_mm: double
child 8, slot: string
child 9, slot_index: int64
child 10, track_id: string
slot_spacing: struct<randomised: bool, reason: null>
child 0, randomised: bool
child 1, reason: null
pick_order: list<item: int64>
child 0, item: int64
arrangement_index: int64
labelling_reason: null
set_index: int64
overlaps: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
topmost_instrument_id: int64
unit_index: int64
arrangement: list<item: int64>
child 0, item: int64
seed: int64
files: struct<depth: string, detections: string, folder: string, image: string, metadata: string, motion: s (... 46 chars omitted)
child 0, depth: string
child 1, detections: string
child 2, folder: string
child 3, image: string
child 4, metadata: string
child 5, motion: string
child 6, videoPrefix: string
child 7, video_dir: string
labelled: bool
slots: list<item: struct<angle_deg: int64, holding_slot: string, instrument_id: int64, position_mm: struct< (... 63 chars omitted)
child 0, item: struct<angle_deg: int64, holding_slot: string, instrument_id: int64, position_mm: struct<x_mm: doubl (... 51 chars omitted)
child 0, angle_deg: int64
child 1, holding_slot: string
child 2, instrument_id: int64
child 3, position_mm: struct<x_mm: double, y_mm: double>
child 0, x_mm: double
child 1, y_mm: double
child 4, slot: string
child 5, stack_index: int64
step_deg: int64
to
{'arrangement': List(Value('int64')), 'arrangement_index': Value('int64'), 'files': {'depth': Value('string'), 'detections': Value('string'), 'folder': Value('string'), 'image': Value('string'), 'metadata': Value('string'), 'motion': Value('string'), 'videoPrefix': Value('string'), 'video_dir': Value('string')}, 'labelled': Value('bool'), 'labelling_reason': Value('null'), 'labels': List({'bbox_xywh': List(Value('float64')), 'centroid_px': List(Value('float64')), 'class_name': Value('string'), 'confidence': Value('float64'), 'has_mask': Value('bool'), 'instrument_id': Value('int64'), 'margin_mm': Value('float64'), 'match_mm': Value('float64'), 'slot': Value('string'), 'slot_index': Value('int64'), 'track_id': Value('string')}), 'overlaps': List(List(Value('int64'))), 'pick_order': List(Value('int64')), 'schema_version': Value('string'), 'seed': Value('int64'), 'set_index': Value('int64'), 'slot_spacing': {'randomised': Value('bool'), 'reason': Value('null')}, 'slots': List({'angle_deg': Value('int64'), 'holding_slot': Value('string'), 'instrument_id': Value('int64'), 'position_mm': {'x_mm': Value('float64'), 'y_mm': Value('float64')}, 'slot': Value('string'), 'stack_index': Value('int64')}), 'step_deg': Value('int64'), 'topmost_instrument_id': Value('int64'), 'unit_index': Value('int64'), 'unlabelled_detections': List(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
arm_pose: list<item: double>
child 0, item: double
depth_intrinsics: struct<d: list<item: double>, distortion_model: string, frame_id: string, height: int64, k: list<ite (... 48 chars omitted)
child 0, d: list<item: double>
child 0, item: double
child 1, distortion_model: string
child 2, frame_id: string
child 3, height: int64
child 4, k: list<item: double>
child 0, item: double
child 5, p: list<item: double>
child 0, item: double
child 6, width: int64
depth_stamp_ns: int64
depth_units: string
detections: struct<camera_frame: string, detections: list<item: struct<axis_direction_valid: bool, axis_elongati (... 1546 chars omitted)
child 0, camera_frame: string
child 1, detections: list<item: struct<axis_direction_valid: bool, axis_elongation: double, axis_end_width_ratio: double, (... 1316 chars omitted)
child 0, item: struct<axis_direction_valid: bool, axis_elongation: double, axis_end_width_ratio: double, axis_endpo (... 1304 chars omitted)
child 0, axis_direction_valid: bool
child 1, axis_elongation: double
child 2, axis_end_width_ratio: double
child 3, axis_endpoints_base_m: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
child 4, axis_length_m: double
child 5, axis_mask_pixels: int64
child 6, axis_negative_end_width_px: double
child 7, axis_positive_end_width_px: double
c
...
d 7, match_mm: double
child 8, slot: string
child 9, slot_index: int64
child 10, track_id: string
slot_spacing: struct<randomised: bool, reason: null>
child 0, randomised: bool
child 1, reason: null
pick_order: list<item: int64>
child 0, item: int64
arrangement_index: int64
labelling_reason: null
set_index: int64
overlaps: list<item: list<item: int64>>
child 0, item: list<item: int64>
child 0, item: int64
topmost_instrument_id: int64
unit_index: int64
arrangement: list<item: int64>
child 0, item: int64
seed: int64
files: struct<depth: string, detections: string, folder: string, image: string, metadata: string, motion: s (... 46 chars omitted)
child 0, depth: string
child 1, detections: string
child 2, folder: string
child 3, image: string
child 4, metadata: string
child 5, motion: string
child 6, videoPrefix: string
child 7, video_dir: string
labelled: bool
slots: list<item: struct<angle_deg: int64, holding_slot: string, instrument_id: int64, position_mm: struct< (... 63 chars omitted)
child 0, item: struct<angle_deg: int64, holding_slot: string, instrument_id: int64, position_mm: struct<x_mm: doubl (... 51 chars omitted)
child 0, angle_deg: int64
child 1, holding_slot: string
child 2, instrument_id: int64
child 3, position_mm: struct<x_mm: double, y_mm: double>
child 0, x_mm: double
child 1, y_mm: double
child 4, slot: string
child 5, stack_index: int64
step_deg: int64
to
{'arrangement': List(Value('int64')), 'arrangement_index': Value('int64'), 'files': {'depth': Value('string'), 'detections': Value('string'), 'folder': Value('string'), 'image': Value('string'), 'metadata': Value('string'), 'motion': Value('string'), 'videoPrefix': Value('string'), 'video_dir': Value('string')}, 'labelled': Value('bool'), 'labelling_reason': Value('null'), 'labels': List({'bbox_xywh': List(Value('float64')), 'centroid_px': List(Value('float64')), 'class_name': Value('string'), 'confidence': Value('float64'), 'has_mask': Value('bool'), 'instrument_id': Value('int64'), 'margin_mm': Value('float64'), 'match_mm': Value('float64'), 'slot': Value('string'), 'slot_index': Value('int64'), 'track_id': Value('string')}), 'overlaps': List(List(Value('int64'))), 'pick_order': List(Value('int64')), 'schema_version': Value('string'), 'seed': Value('int64'), 'set_index': Value('int64'), 'slot_spacing': {'randomised': Value('bool'), 'reason': Value('null')}, 'slots': List({'angle_deg': Value('int64'), 'holding_slot': Value('string'), 'instrument_id': Value('int64'), 'position_mm': {'x_mm': Value('float64'), 'y_mm': Value('float64')}, 'slot': Value('string'), 'stack_index': Value('int64')}), 'step_deg': Value('int64'), 'topmost_instrument_id': Value('int64'), 'unit_index': Value('int64'), 'unlabelled_detections': List(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.
IRIS VLM Instrument Arrangement Dataset
Observations of five surgical instruments laid out in a row by a UFACTORY 850
arm, each paired with a video and a motion log of the arm clearing that layout
back into its holding fixtures. Collected by the VLM Data Collection routine
in the IRIS dashboard.
93 cycles, one folder each, 2.6 GB.
What one cycle is
The routine repeats a two stage cycle:
- Lay out. The arm takes the five instruments from their holding fixtures and places them in a row, in the order the cycle's arrangement names, each at its own yaw angle and its own X spacing. Placements happen in a drawn order, so instruments overlap and stack differently every time.
- Observe and clear. The arm returns to the survey pose, one colour frame and one depth frame are captured, the detector is run over the scene, and then the arm picks the row apart, topmost instrument first, while the wrist camera records.
Files
Each cycle is a folder u<NNNN>_arr<slot order>, for example
u0007_arr3-1-4-5-2, holding six files:
| File | What it is |
|---|---|
u<NNNN>_arr<...>.png |
Colour frame of the arrangement, from the survey pose |
u<NNNN>_arr<...>_depth.png |
Aligned depth frame from the same moment (16-bit) |
u<NNNN>_arr<...>.json |
The routine's own sidecar: arrangement, per-slot ground truth, labels |
u<NNNN>_arr<...>_detections.json |
Raw detector output: masks, boxes, intrinsics, arm pose |
u<NNNN>_iris_color_<timestamp>.mp4 |
Video of the picks that cleared the row (~60 s each) |
u<NNNN>_motion.json |
Every arm and gripper command made while the video was recording |
Sidecar (schema_version: iris.vlm.observation.v1)
arrangement— which instrument sits in each of the five slots, left to right.slots[]— per slot ground truth:instrument_id,angle_deg(yaw the arm placed it at),position_mm(the X/Y it was actually placed at, not the taught slot),stack_index(placement order, so higher sits on top),holding_slot(the fixture it came from).labels[]— the detector's instances matched to slots:class_name,confidence,bbox_xywh,centroid_px,has_mask, plusslot,instrument_id,match_mm(how far the detection fell from the slot) andmargin_mm(how much closer that slot was than the next one).labelled/labelling_reason— whether all five instruments matched a slot, and why not when they did not.overlaps,topmost_instrument_id,pick_order— the stacking the routine built and the order it took the row apart in.seed,arrangement_index,set_index,unit_index,step_deg,slot_spacing— how the cycle was drawn.
Motion log (schema_version: iris.motion.log.v1)
recording_started_at / recording_stopped_at in bridge host unix seconds,
then steps[]: one entry per arm or gripper command, with its duration. Poses
are [mm, mm, mm, rad, rad, rad], joints are radians. 56 steps per cycle.
Aligning steps to video frames means offsetting a step's timestamp against
recording_started_at.
Instrument classes
scalpel_handle_large, scalpel_handle_small, surgical_scissors,
tissue_forceps, artery_forceps.
Contents
- 93 cycles, 24 distinct arrangements, 4 sets of each (the last arrangement is cut short at 1 — the run was stopped mid-collection).
- One seed for the whole run:
817842836. Yaw step 15°. - 90 cycles have randomised slot spacing; 3 fell back to the taught slots.
- Videos are ~60 s each, 2.5 GB of the 2.6 GB total.
Known limits
Read these before training on it.
- 29 of 93 cycles are not fully labelled (
labelled: false). In those, the detector matched only 3 or 4 of the 5 instruments to a slot, usually because confidence fell under the 0.50 threshold. Theslots[]ground truth is still complete and correct for every cycle; it is the detector-derivedlabels[]that is short. Filter onlabelledif you need complete mask supervision. - The videos start late. Recording was started and the first pick began
immediately after, but the camera needs a few seconds to actually be filming,
so the opening of each video can miss the start of the first pick. Every video
in this dataset was collected before that was fixed. The motion log covers the
full pick phase regardless, so a step near
recording_started_atmay have no frames behind it. - The run was stopped part way through its schedule, so arrangement coverage is uneven: 23 arrangements have 4 sets, one has 1.
- All frames come from one camera at one survey pose, on one table, under one lighting setup. No viewpoint or scene diversity.
- Depth is 16-bit PNG; the units and intrinsics are in each cycle's
_detections.json(depth_units,depth_intrinsics), not in the PNG.
Hardware
UFACTORY 850 (XI15) arm, Intel RealSense depth camera, ROS 2 bridge. Captured 2026-09-28.
- Downloads last month
- 375