viola_cube_pick_v1

50 teleoperated demonstrations of a StarAI Viola arm picking up one red 30 mm cube from a random spot on a desk and dropping it into a glass. A front RGB-D camera and a wrist camera at 30 fps, follower joint states and leader-arm actions, recorded with LeRobot 0.6.2 in the LeRobot v3.0 format. The cube ends up in the glass in all 50 episodes.

The setup is deliberately small: one arm on a desk, one cube, one glass. It was recorded to train an ACT policy and test it on the same rig.

Robot	StarAI Viola follower arm (6 joints + gripper), driven by a StarAI Violin leader arm
Episodes	50, recorded in 5 sessions between 2026-10-06 and 2026-10-11
Frames	37,504 at 30 fps (20.8 minutes)
Cameras	up: Orbbec Gemini 2 in front of the arm, colour and colourised depth stacked in one 880 x 640 image. wrist: UGREEN USB webcam on the gripper, 640 x 480
State and action	7 floats each: 6 joints + gripper, in each arm's calibrated range
Task string	Pick up the cube and place it in the target zone (the target zone is the glass)
Format	LeRobot v3.0 (codebase_version: v3.0), AV1 video in MP4, 1.24 GB
License	CC BY 4.0

Task

Before each episode the cube is placed by hand at a random spot and orientation on the desk in front of the arm, inside the area the cameras see. The operator then moves the leader arm and the follower copies it: approach, grasp, lift, carry to the glass, release over it, and bring the arm back to a resting pose, where it stays until the episode's time cap. The glass stands at the same spot on the right of the workspace in every episode. One cube (red, 30 mm, 18 g) and one task string for the whole dataset.

Hardware

• Follower: StarAI Viola, 6+1 degrees of freedom, made by Fashion Star and documented for LeRobot on the Seeed Studio wiki. LeRobot plugin lerobot_robot_viola 0.0.4 (robot_type: starai_viola).
• Leader: StarAI Violin (same kinematics), plugin lerobot_teleoperator_violin 0.0.4.
• up camera: Orbbec Gemini 2 on a tripod facing the arm across the desk. Colour 640 x 480, depth 640 x 400, both at 30 fps. Its view did not change across the 50 episodes.
• wrist camera: UGREEN USB webcam clipped to the gripper, looking along the fingers; 640 x 480 MJPG, auto exposure.
• Recording loop: LeRobot's lerobot-record through a small wrapper that holds 30 Hz and limits each follower joint to 5 units per tick (150 units/s). In rig tests the follower trailed the leader by 100-170 ms.

Data format

LeRobot v3.0 stores all episodes back to back in shared files; meta/episodes/ says where each episode starts and ends.

data/chunk-000/file-000.parquet ... file-005.parquet            joint states, actions, indices
meta/info.json, meta/stats.json, meta/tasks.parquet
meta/episodes/chunk-000/file-000.parquet ... file-003.parquet   per-episode offsets and stats
videos/observation.images.up/chunk-000/file-000.mp4 ... file-009.mp4
videos/observation.images.wrist/chunk-000/file-000.mp4 ... file-005.mp4

Feature	Type	Content
observation.state	float32 [7]	follower joint readings: Motor_0.pos ... Motor_5.pos, gripper.pos
action	float32 [7]	leader joint readings at the same tick, same names
observation.images.up	video, 880 x 640 x 3	front camera: colour on top, colourised depth below (see below)
observation.images.wrist	video, 480 x 640 x 3	wrist camera
timestamp, frame_index, episode_index, index, task_index		LeRobot bookkeeping

Videos are AV1 (SVT-AV1, CRF 30, a keyframe every 2 frames), yuv420p, 30 fps.

Joints and units

Each joint value is LeRobot's normalised position: -100..100 spans the joint's calibrated range on that arm, and the gripper runs 0..100 with 0 = closed. On the follower the calibrated ranges are:

Name	Joint	-100..+100 on the follower (degrees)
Motor_0	base	-122.1 .. +123.0
Motor_1	shoulder	-88.1 .. +70.9
Motor_2	elbow	-100.0 .. +91.0
Motor_3	wrist roll	-180.0 .. +180.0
Motor_4	wrist pitch	-109.1 .. +94.9
Motor_5	wrist rotation	-176.0 .. +176.0
gripper	gripper	0 = closed .. 100 = open

The leader was calibrated separately, so the same number on the two arms means nearly the same angle. Parked at the end of an episode, leader and follower read within about 3 units of each other on every arm joint, usually within 1.

Gripper: the follower's fingers meet at about 18, and a held 30 mm cube reads about 52 (roughly 0.8 mm of opening per unit). In this dataset the follower's gripper never reads below 45.8 because it always closes on the cube, while the leader's gripper reading (action) goes down to 0.

The up image: colour and depth in one frame

Each up frame is two images stacked vertically, both 640 pixels wide:

• Rows 0-479: the colour image (RGB).
• Rows 480-879: the depth image, colourised by the camera driver before recording. Depth readings between 100 and 600 mm are min-max normalised over that frame's own pixels in the window, then mapped through OpenCV's JET colour map. Pixels outside the window, and pixels without a depth reading, are black.

What that means in practice:

• The colours are not metric. The normalisation is per frame, so a colour only says near or far relative to the rest of that frame. The same desk spot changes colour when the arm enters the view and moves the frame's nearest point. Use the depth half as a shape and occlusion cue. The raw depth values were not stored.
• Red is near, blue is far. The driver writes the JET map in OpenCV's BGR channel order into an RGB frame, which flips the usual reading: in the stored image the nearest in-window surfaces are red and the farthest are blue.
• The halves are not aligned. Colour and depth come from separate sensors with different fields of view and are stacked without registration, so pixel (r, c) in the colour half and pixel (r + 480, c) in the depth half show different points.
• info.json marks the feature as a plain RGB video (is_depth_map: false). A policy that takes the whole 880 x 640 frame sees both halves as one image.

The conversion is turn_depth_to_color() in the Orbbec camera integration used with LeRobot (lerobot/cameras/orbbec/camera_orbbec.py, from github.com/ZhuYaoHui1998/orbbec at commit 83be839).

How it was collected

Session	Date	Episodes	Light	Episode cap	Reset between episodes
1	2026-10-06	0-9	daylight	20 s	10 s
2	2026-10-07	10-19	evening, desk lamp	30 s	10 s
3	2026-10-08	20-29	evening, desk lamp	25 s	5 s
4	2026-10-11	30-39	evening, desk lamp	25 s	5 s
5	2026-10-11	40-49	evening, desk lamp	25 s	5 s

• One operator, one camera setup and one calibration for the whole dataset.

• Every episode runs to its time cap (599-601, 897-901 or 744-751 frames). After the drop the operator brings the arm back to a resting pose and waits (the raised pose below in session 1, folded from session 2 on). Each episode starts with the arm still for 1.6-4.6 s (median 3.1 s) and ends with it still for up to 16.6 s (median 8.6 s).

• Start poses (first frame of observation.state, in the units above):

  Episodes	Start pose
  0, 10, 20, 30 (first of a session)	the vendor connect pose (0, -100, 61, -1, 31, 0, 50), right after the arms connect
  1-9	a raised pose with the wrist camera looking down at the desk; median (7, -20, 18, -7, 42, 10, 80)
  11-19	folded, between the vendor pose and the rest pose; median (5, -100, 80, -2, 32, -1, 76)
  21-29, 31-49	near the folded rest pose (0, -100, 97, 1, 1, 0, 50); median (2, -100, 98, 0, 10, -1, 72)

  The arm was brought back by hand, so the elbow, wrist pitch and gripper vary by up to about 40 units between episodes of the same group.

• One take replaced. In session 3 one episode was deleted because at its start the cube was in neither the up colour image nor the wrist image. The later episodes were renumbered, and episode 29 was recorded the same evening as its replacement. Episode indices run 0-49 without gaps.

• No cube coordinates were recorded. The spots cover the part of the desk the cameras see.

Known quirks

• The operator's hand is in view for a moment in some episodes. In 16 episodes a hand passes through the top-left corner or along the left edge of the up view for 0.1-0.6 s, mostly while the arm stands still before or after the task. It is there for the first 0.4 s of episode 40, and in episode 49 a forearm and hand cross the left third of the up view for the first 0.6 s and the top of the wrist view for the first 0.2 s.
• Cube at or below the bottom edge of the colour view at the start of several episodes, among them 4, 6, 8, 10, 13, 14, 16, 21, 24, 25, 33, 35, 38, 39, 43, 47, 48 and 49. The cube is still visible in the depth half and the wrist view.
• action is the leader's raw reading. The follower received a rate-limited version of it, so a quick flick it could not follow appears in action and never in observation.state.
• The cameras are not hardware-synchronised. The two images of one frame can be up to about one frame interval (33 ms) apart; the offset was not measured.
• The wrist webcam runs on its own clock at about 28.7 fps. While the arm moves, 4-4.5 % of the wrist frames repeat the one before (two episodes checked).
• Dark wrist images in the evening sessions. Under the desk lamp the wrist view is much darker than in the daylight session.
• Servo read stalls. When a servo reply is missed the read waits up to 100 ms, so the joint readings hold still for a few frames now and then. The logs of sessions 1, 2 and 4 counted 1, 4 and 9 such stalls.
• Idle stretches at the start and end of every episode (see above). Trimming them is a reasonable preprocessing step for some policies.

Intended use

Imitation learning for one-arm pick-and-place with a front and a wrist camera, for example ACT or Diffusion Policy in LeRobot.

An ACT policy trained on all 50 episodes (LeRobot 0.6.2, ResNet-18 backbone, chunk size 100, 100,000 steps at batch size 8) was run on the same rig in a small A/B on 2026-10-14, with the cube on 5 marked spots and 30 s per run. Started from a hand-set "pounce" pose (-0.4, -88.6, 63.4, -0.6, 45.1, -0.6, 48.6), close to the vendor pose, it put the cube in the glass in 7 of 10 runs. Started from the folded rest pose, it did not move in any of 5 runs. Picks were counted by eye at the rig. That is 15 runs in one evening, a smoke test of the data.

Limitations

• Small: 50 episodes, 21 minutes, one operator.
• One object, one colour, one target at one place, one room and one camera placement. Expect a policy trained on it to need new data for a different object, target, room or camera pose.
• Lighting is uneven: 10 episodes in daylight, 40 under an evening desk lamp.
• All 50 episodes succeed; there are no failed attempts to learn from.
• The start pose depends on the session (see above), and the A/B above shows that a policy trained on this data is sensitive to where it starts.
• The depth half is a per-frame colour map, not metric depth.
• Joint units come from these two arms' calibrations. Another Viola calibrates to slightly different ranges, so the numbers will not transfer one to one.

Loading

Recorded and tested with LeRobot 0.6.2; LeRobot 0.4.4 reads it too (both use the v3.0 dataset format).

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("aitron7/viola_cube_pick_v1")
frame = ds[0]
up = frame["observation.images.up"]           # torch float32, 3 x 880 x 640, values 0..1
colour, depth_vis = up[:, :480], up[:, 480:]  # colour image, per-frame depth colour map
state, action = frame["observation.state"], frame["action"]

Watch an episode locally with Rerun:

lerobot-dataset-viz --repo-id aitron7/viola_cube_pick_v1 --episode-index 0

Citation

@misc{viola_cube_pick_v1,
  title        = {viola\_cube\_pick\_v1: StarAI Viola cube pick-and-place demonstrations},
  author       = {aitron7},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/aitron7/viola_cube_pick_v1}}
}