Contact Us
Robotic Arm Manipulation

A dual-arm platform on a wheeled base — in a kitchen, an office pantry and on a parcel line — with every sub-action attributed to the arm that performs it.

5 Tasks · 7 Episodes

One bimanual handover-and-scoop, three single-arm tasks, and three clips of both arms clearing a parcel conveyor in parallel.

378 Primitives

Every episode annotated to the 14-primitive standard, each row carrying a contact state and the signal its boundary came from.

250 Hz · 6 Streams

Teleoperated episodes: both arms at 250 Hz, a head camera and one per wrist, RGB and depth. The parcel-line clips are lighter — three RGB cameras and both arms at 30 Hz.

BEHIND THIS SAMPLE
~2,500task-hours

The seven episodes below are a sample. The collection behind them runs to roughly two and a half thousand task-hours on the same platform, across eight scene families — weighted towards the home and the office, with a long tail through retail, food preparation, manufacturing and clinical work.

28%
22%
14%
12%
8%
8%
01 700 h · 28%
Home & Domestic

Living room, bedroom, study, kitchen, dining room, balcony, courtyard and entryway. Storage, bathroom, vanity, laundry and household cleaning. Children's activity tables.

02 550 h · 22%
Office & Focused Work

Workstations, meeting rooms and collaboration areas. Desks and workbenches, and the shared support space around them.

03 350 h · 14%
Retail & Customer Service

Supermarkets, shops and in-store shelving. Reception, checkout, entrance and customer-service points. Bars and drinks counters, dining and lounge seating.

04 300 h · 12%
Food Preparation

Prep stations, fresh-food processing and back-of-house kitchens. Work surfaces and island counters, service passes through to the dining floor.

05 200 h · 8%
Warehouse & Back-of-House

Goods-in and receiving, circulation routes, technical and storage rooms, back-of-house areas and the corridors between them.

06 200 h · 8%
Manufacturing & Assembly

Electronic components, appliances and consumer electronics. Machining, plastics, garment work and handcraft. Assembly benches.

07 125 h · 5%
General & Unscoped

Simple and unscoped scenes, general everyday operations, and mixed-use areas such as the washbasin and mirror cabinet that sit between home and cleaning.

08 75 h · 3%
Clinical & Sanitation

Operating theatres, sterilisation rooms and sanitation areas.

Hour figures are indicative of the shape of the collection, not an audited breakdown. The sampled episodes below are drawn from families 01, 04 and 05.
TASK
PARCEL LINEpolicy rollout · 10 s clips · both arms
0.00 / 0.00
STREAMS —
LEFT ARM—
· PRIMITIVES—
RIGHT ARM—
· PRIMITIVES—
SCOOP CYCLE
HANDOVER
—
PRE_PICK
PICK
TRANSPORT
PLACE_AND_RETURN
DECLARED EXTENSION
—
GRIPPERopening · both arms
Verify the arm assignment yourself
The delivered annotation and the delivered streams disagree about which arm does what — but not consistently. Two of these four episodes name the wrong arm and two are correct, so a global correction would break as much as it fixed and every episode has to be checked on its own. The roles here are read off the motion. On a single-arm task the test is blunt: one arm never opens its gripper and never leaves its parking pose. On the bimanual task, measure how far each end-effector travels between the two handovers and the tool arm separates from the anchor arm by an order of magnitude.
travel[arm] = Σ ‖p[i+1] − p[i]‖₂ # ee_positions, whole episode width[arm] = max(gripper) − min(gripper) inert = travel < 0.30 m and width < 5 mm # single-arm: this arm is parked tool = argmax(travel over the middle window) # bimanual: sweeps source → target anchor = argmin(travel over the middle window) # bimanual: clamped to the work # the cross-check that makes the bimanual case trustworthy: d(t) = ‖ T_left⁻¹·p_left(t) − T_right⁻¹·p_right(t) ‖₂
Two independent extrinsics push the two end-effectors into one camera frame. At each handover they land 2.3 cm apart, and over 40 cm apart the rest of the time. Calibration wrong enough to flip left and right could not produce that.
episodedata sayslabels sayevidence
Cereal · two armsleft = toolright = tool3.84 m vs 0.25 m travel
Salt shakerright actsleft actsleft gripper never opens
Matcharight actsright actsagrees
Water bottleleft actsleft actsagrees
Parcels · clips 03 / 05 / 08both actprompt names left only4 + 4 grip cycles per clip
On the parcel line the question is not which arm but whether both do: each clip carries four closing and four opening ramps on each gripper, the left wrist camera looks into the pile whenever the left fingers are shut and the right one looks onto the outbound belt whenever the right fingers open. The vendor prompt describes only the left arm's half.
On the bimanual episode the separation is 15×, and it agrees with what the wrist cameras show — the right camera is nearly a still image for forty seconds while the left one sweeps. On the salt-shaker episode the left gripper never moves at all: 0.0 mm of width change and 3 cm of travel across the whole take, against 2.09 m for the right.
Every span, boundary and cycle count on this page comes out of that same derivation, and each carries the evidence that produced it in the panel above.