Training data from the physical world.
Collov Data turns real spaces and real robots into datasets physical AI teams can train on today — parsed interiors, spatial evaluation sets and 2.3 million real-robot episodes.
Unretouched first frames from delivered episodes. Every frame maps to a specific episode, robot serial number, capture configuration and MCAP recording you can replay before you buy.
Explore our other data business
Looking for expert training and evaluation data?
Explore Coding Agent, General Agent and Reasoning datasets in Expert Data.
Spatial data
The spatial data engine.
Every run of Collov's visual agent produces structured spatial data as a by-product. Collov Data packages it for frontier labs, robotics teams and enterprises whose AI has to work indoors — with people reviewing along the way.
The loop that runs Collov's products — perceive the scene, plan, act, check the result, learn — writes a structured record at every pass. That record is the dataset.
Parsed scenes
Real interiors broken down into objects, materials, geometry and the spatial relationships between them — structured, queryable scene state rather than flat labels.
Edit trajectories
Complete multi-step edit histories from production traffic: what was asked, what changed, what was corrected. Not synthetic one-shot before/after pairs.
Spatial evals
Benchmarks that check whether a model respects real-world constraints — depth, planes, occlusion, scale and placement — built on the same approach as Collov's Spatial Arena.
Humans stay in the loop: reviewers check and correct agent output before it becomes training data.
Physical AI data
Real-robot manipulation data at foundation-model scale.
A production data supply, not a research dump. Licensed, deduplicated and physics-verified episodes captured on real hardware, shipped as MCAP and LeRobot v3.0 — and expandable to your task list on a six-week capture cycle.
Why internet video can't train a robot.
Compute scaled. Architectures scaled. Physical interaction data did not. Passive footage is missing the four things a manipulation policy needs most.
No contact forces
Video shows what moved, never the force that moved it. Grasp stability can't be learned from pixels alone.
No action labels
No joint targets, end-effector commands or gripper state — nothing for a policy to imitate.
No physical consistency
Annotators can't check momentum or kinematic limits, so noise and teleop artefacts slip into training.
No embodiment transfer
Single-robot corpora overfit one kinematic chain and fail on a new body.
We record the full causal loop.
The robot's own state at up to 1 kHz, the commanded action, the resulting contact and the sensed outcome — time-synchronised in one MCAP episode.
- 1State
- 2Intervention
- 3Outcome
- 4Feedback
Eight commercial domains, collected to production depth.
Every domain below is live, delivered volume — not a roadmap. Select one to see its scenarios.
Clothing foldingHome service robotics
The deepest domestic manipulation corpus in the catalog: kitchens, bathrooms, bedrooms, living rooms and entryways, captured in real furnished apartments rather than staged rigs.
Delivered episodes by scenario
Top 17 of 250 scenarios. 233 more in the library, including kitchen cleaning, cooktop cleaning, mixed-clutter sorting and food handling.
Thirty-four platforms from eleven OEMs.
Wheeled mobile manipulators, full-size humanoids and force-controlled arms — the range of bodies a generalist policy needs, in one licence.
Delivered hours, top seven platforms
Fleet composition
Parallel grippers, five-finger dexterous hands and vacuum tooling, with 6-axis wrist force/torque sensing.
Signal density
Not a video with a caption. A multi-stream recording.
Every episode is a time-synchronised, schema-typed recording you can replay topic by topic — vision, action and force aligned to one episode clock.
Native sample rates, as recorded
Rates are as recorded on production lines, not resampled.
400+ recorded topic types
RGB and depth
Head, chest, torso, wrist and gripper cameras.
Proprioception
Joint position, velocity, current and command.
Force and torque
6-axis wrist F/T and end-effector wrench.
Frames and metadata
TF tree, camera intrinsics and extrinsics, SOP.
Labelled to the step, not the clip.
Every task breaks down into an ordered sequence of steps, and every step carries skill verbs from a controlled 140-term taxonomy.
Worked example: kitchen cleanup and dish storage
Unitree G1, bimanual, stereo and monocular 720p, BrainCo hands. 22 annotated steps and 13 skill verbs in total; first eight shown.
- Move the bipedal robot to the sinkMove
- Grasp the kettle with the right armGrasp
- Carry and place the kettle at its home positionCarry, Place
- Bend and collect used tissues from the counterBend, Grasp
- Wipe standing water from the work surfaceWipe
- Open the dishwasher doorOpen
- Place dishes into the rack slots, repeating per itemPlace, Repeat
- Close the door and return to the home poseClose, Adjust
From the 140-verb skill taxonomy
Human egocentric capture, as a complement.
Wearable first-person recordings of real human work — the cheapest route to task diversity. 40 scenarios, 1,400 recorded hours, nine rigs in rotation.
Household
Washstand cleaning, cucumber cutting, dish washing, after-meal tidying, wok stir-frying.
Factory
Boxing, weighing, gluing and sealing at a packing station; manual pallet-jack transport.
Hotel
Guest-room housekeeping: linen changes, fixture wiping, amenity arrangement.
Office and outdoor
Restroom cleaning in stereo; a six-camera outdoor waste-station rig.
RoboPocket
Compact chest or head monocular rig for long household sessions.
MRR
Chest-mounted stereo pair with dual-video output per session.
WE-Ego
Wide-FOV fisheye headset for cluttered interior work.
DAS-Ego
Six-camera cart-mounted array for outdoor and multi-worker scenes.
Our inverse-dynamics model recovers 6-DoF end-effector trajectories and contact points from this footage, turning passive video into robot-executable action labels.
Physics verification, not human eyeballs.
The EdgeWAM engine scores every trajectory against rigid-body physics in both directions before it enters a delivery batch.
Forward dynamics: consistency scoring
Does this episode obey physics?
Trajectories are checked against rigid-body kinematics, collision boundaries, joint limits and acceleration and jerk envelopes. Sensor dropouts, tracking jumps and teleoperator noise are flagged and quarantined before training — not discovered three epochs into your run.
Inverse dynamics: latent action recovery
What action produced this frame?
Recovers 6-DoF end-effector trajectory distributions and contact points directly from passive egocentric video, converting unlabelled archives into structured, robot-executable actions at a fraction of teleop cost.
| Metric | Benchmark target | Validated against | Status |
|---|---|---|---|
| Kinematic trajectory accuracy | MAE ≤ 3.1 mm / ≤ 2.8° | OptiTrack optical ground truth | Verified |
| Physical anomaly detection | 94.2% precision / 91.8% recall | Kinematic-limit and jerk injection | Verified |
| Contact force estimation | MAE ≤ 1.2 N (0.5–10 N window) | 6-axis F/T load-cell fixtures | Verified |
| Downstream sample efficiency | 38% fewer training hours | Policy convergence vs. raw data | Verified |
What it does to your policy.
Same architecture, same compute budget, same task suite. Only the training corpus changes.
Task success rate after convergence
over unfiltered raw video. Physics-verified tokens remove the non-physical demonstrations a policy would otherwise learn to imitate — the largest single source of brittle failure at deployment.
Delivery and licensing
Lands in your pipeline in the format it already reads.
Slice the robot catalog by platform, domain, scenario, skill or capture configuration — or take a pre-cut dataset off the shelf.
Example dataset card
Unitree G1 diverse manipulation
250 tasks across 90 home and retail scenarios — cleaning, tidying, clothing handling, refrigerator zoning, basket shopping and shelf restocking. Six capture configurations span Dex1 and BrainCo end effectors with monocular and stereo 720p vision.
Largest scenarios
Formats
- MCAP — lossless multi-stream, schema-typed, replayable
- LeRobot v3.0 — drop-in for Hugging Face training loops
- OWAMcap — world-model token stream
- Parquet or webdataset shards on request
Slice by
- Robot platform, brand or end effector
- Domain, scenario, task or skill verb
- Capture configuration: mono, stereo, RGB-D
- Duration, success flag or object class
Ingestion
Standardised capture in MCAP and LeRobot v3.0, with device serial, capture config and camera calibration bound to every record.
Sanitisation
Automated GDPR/CCPA facial and biometric scrubbing, with signed commercial waivers from every operator and location.
Verification
EdgeWAM physics scoring, duplicate detection and a causal metadata graph for every episode.
Deployment
S3/GCS transfer or physical shipment under a tier-1 SLA. You pay only for frames that clear the agreed thresholds.
Every licence includes
- Perpetual, worldwide commercial training rights
- Full provenance chain: operator waiver, robot serial, episode
- A personal-data scrub certificate with each delivery batch
- A replayable MCAP sample set before any commitment
New capture, on your task list
Monthly robot-hours of net-new capture
Today's run-rate by line
How a commission runs — first verified delivery in six weeks
Take the sample set, then take the catalog.
Tell us the embodiment, the skills or the model capability you're training. We'll cut a free evaluation slice against your task list — full annotation, no commitment.
