Unstaged · In-the-Wild · PII-Protected · Ready to License

Real behavior,
recorded where
it happens

Human behavior data for Physical AI, captured from real customers in real stores — nobody recruited, nothing staged.

Manipulation · hand–object
ARCADECheckout

736,560 episodes of hand–object interaction a year, from one self-checkout lane — 2,728 hours. Real customers scanning their own items, recorded from fixed multi-view cameras. Grasp, transport, scan and retry differ from person to person, and failures and recoveries occur at their natural frequency.

736,560 episodes / yr · per lane
2,728 h / yr · per lane
10 lanes per store
15,000–30,000 SKUs

A recorded lane session becomes episodes here: grasp and release are marked per hand, the stream is cut at those boundaries, and every frame carries a measured 3D hand and object pose.

Unstaged in-the-wild manipulation data Scan success and failure signals are aligned with the placing and re-scanning that follow, on one timeline.

Five fixed exocentric cameras per lane, hardware-synchronized. On-device AI extracts 3D human pose (hands included), 6-DoF object pose, action-segment labels and scanner success/fail logs on site, all aligned on a single timeline. SKU and POS events linked under pilot agreements. Commercial license; sample subsets available through gated access on request.
Request sample
Capture
Fixed exo multi-view · 5 cams per lane · hardware-synced
Volume
736,560 episodes · 2,728 h per lane per year · 2,160 episodes per operating day
Environment
10 self-checkout lanes per store · 15,000–30,000 SKUs in circulation · continuous collection, not a one-off shoot
Subjects
Real shoppers, not recruited · no raw video in the dataset · no audio captured
Formats
Pose JSON / NPZ · Episode Parquet · COCO-style labels
Retargeting
Robot model import → motion retargeting → NVIDIA Isaac training setup (optional)
License
Commercial license · no redistribution · academic terms on request
Taxonomy · 01

Action taxonomy across 10+ classes Self-checkout manipulation classified into more than 10 action classes plus failure–recovery branches. Percentages are the share of episodes containing each action.

Not a list designed for demonstration — behaviors were observed first, then classified.

Approach100%

ApproachHand enters toward the item in basket or cart

Grasp100%

GraspOne-hand, two-hand, re-grasp included

Lift98%

LiftTrajectories by item weight and shape

Hand-over23%

Hand-overItem passed from one hand to the other mid-motion

Transport97%

TransportPath and speed to the scanner

Scan100%

ScanBarcode presentation pose and angle

Rotate31%

RotateSearching for barcode orientation — branch after failure

Re-grasp14%

Re-graspSlips, two-hand assist and re-alignment

Re-scan27%

Re-scanRetry after a failure signal

Place100%

PlacePlacement onto bagging area or bag

Open set

and countingA class is added when a behavior shows up often enough to label. The list follows the data, not a shot list.

Modalities · 01

Every episode, one timeline Every channel is aligned to the same timestamps.

  • Multi-view time-synced capture5 cams
  • 3D human poseincl. 21 kpt hands
  • 6-DoF object poseper object
  • Action-segment labelsstart · end · class
  • Scan outcomesuccess · fail
  • SKU · POS eventspilot
  • Retargeted robot motionoptional · Isaac
Pipeline

From a frame to a policy

One clip moves through four stages at once, so you can see frame by frame where a motion holds and where it breaks. Each stage hands off a file, not a screenshot.

Egocentric

The robot's own point of view

Third-person capture is where the data starts, not where it ends. Replayed in simulation, one episode also renders the views the robot itself would have had — per embodiment, without mounting a single new camera in a store.

Licensing · Privacy · IP

Clean to license.
Clear from day one. This data is recorded from real customers, so we answer how it was collected and what was removed before anything else.

01

On-device extraction

3D pose and 6-DoF object pose are extracted on site by on-device AI. Raw video never leaves the store and is in no dataset we license, and no audio is captured.

02

Consented sites

Collected only at sites with a data-use agreement and customer notice in place. Store and brand identifiers are anonymized.

03

Commercial license

Commercial license for model training and evaluation. No redistribution or resale. Academic collaboration on separate terms.

04

Gated access

Sample subsets through gated access on request. Full datasets delivered to a dedicated bucket after contract, with versioning and a change log.

Custom data · Design partner

Define the next
batch with us We define the episode spec and failure taxonomy around your target task, embodiment and data structure, then produce the first batch in live stores. It starts with a 30-minute Technical Fit Review.

ARCADECheckout
Episode schema · v0.4 · sample
Delivery layout

    
episode.json

    

Illustrative sample — field names and values are not final