Raw footage in. Training-ready episodes out.
We turn raw robot & egocentric footage into training-ready episodes: object tracks, hand poses, action labels.
Capture footage to spec, in the field.
Equip your own operators or tap the Rimb contributor network. The capture app guides every clip so it clears QA the first time, with consent recorded at the source.
Record clips into a collection, then publish when ready.
Bring footage from anywhere. It lands in one dataset.
Connect a bucket, paste a link, or shoot in the app. Clips are ingested, deduped, and versioned into a single training set.
Object tracks, hand poses, and language. Every layer reviewed.
Drag the scrubber on a real annotated episode below: object tracks, hand poses, and action labels move with every frame. It's the exact format you'd receive.
Best viewed on a desktop browser — here's a recording instead.
What ships in every episode.
Each episode arrives as a stack of reviewed layers, not a raw clip: object tracks, hand poses, action and language labels, plus the proprioceptive signals manipulation policies train on.
Every object detected and tracked frame-to-frame with a stable identity.
Both hands as 21-keypoint poses, grasp by grasp.
Temporal segments marking each action and sub-task in the episode.
Action-aligned natural-language descriptions, anchored to the frames.
Per-episode provenance: who appears, what they agreed to, what rights attach.
Per-frame 8-dim proprioceptive state and action vectors, in RLDS / LeRobot.
Every label reviewed against explicit criteria.
Unreviewed labels add noise your model will memorize. We set the reject criteria up front, then check every episode before it ships.
Export straight into your framework.
The same episode, written to the format your training stack already reads, so there's no glue code to write.
dataset/ ├─ data/*.parquet └─ videos/*.mp4
step {
observation,
action, reward
}episode.mcap ros2 · foxglove ready
The paper trail ships with the episode.
Every episode carries a consent packet: who appears, what they agreed to, where the footage came from, and what rights attach. It's stored as fields in the data, not a separate PDF.
Egocentric footage is footage of people: hands, homes, workplaces. IL BIPA · TX CUBI are long-standing biometric consent law, and CA AB 2013 put training-data transparency duties in effect on Jan 1, 2026. In the EU, the AI Act's Article 10 adds data-governance duties for high-risk systems. Buyers increasingly ask where the footage came from before they buy.
- consent_id
- contributor_ref · pseudonymous
- capture_date · capture_context
- consent_scope · commercial_ml_training
- biometric_release
- rights_assignment
- revocation_contact
We are not lawyers. We do not certify compliance. We structure the documentation so you can.
Any task, any environment.
We capture footage wherever the work happens: kitchens, warehouses, labs. Hover a sample to see footage.
Food service & kitchens
Wet, occluded, deformable objects handled at speed.
Household & domestic
Tidying, laundry, and kitchen chores: long-horizon tasks in cluttered, real homes.
Warehouse & logistics
Pick, pack, and tote handling at fulfillment speed, across a changing mix of SKUs.
Manufacturing & assembly
Contact-rich insertion, fastening, and fixture work that needs frame-accurate labels.
Retail & service
Shelf restocking and front-of-house handling in spaces shared with people.
Healthcare & lab automation
Tube, tray, and instrument handling under protocol, where provenance matters.
Request 5 free annotated episodes.
Describe your robot task below. If it's a fit, we'll annotate 5 episodes on that task and send them back free in 7–10 days. No commitment required.
STEP 1 · FREE
5 annotated episodes
Describe your task below. If it's a fit, we annotate 5 episodes and send them back free in 7–10 days. No commitment.
RLDS · LeRobot v2.1 · MCAP
STEP 2 · AFTER SAMPLE
Full episodes pilot
If the sample confirms fit, we move to a fixed-scope pilot on your task, evaluated against a metric you define in writing before we start.
object tracks · hand poses · language · consent