Humanoid Training Data
Manipulation data from inside a working factory
Either we capture it for you across a contracted network of production plants, or you send us footage you already hold. Both routes return the same structured, labelled, consented datasets.
The network is the product
Processing footage is a solved problem for a competent team. Getting inside a production plant that is willing, prepared and legally clear to share what happens on its floor is not. That is the part we have already built.
Contracted and ready
A large and growing network of production factories already signed up for data sharing, so a request turns into a capture run rather than a search for a pilot site.
Hundreds of stations
Multiple industrial sectors, hundreds of distinct work stations and process types, so a brief can be matched to the work rather than the other way round.
New tasks come online quickly
New plants and new task types are added inside the framework that already exists. There is nothing to renegotiate each time the brief changes.
Variety inside one task
The same task recorded across different plants, operators and part mixes, which is what makes a policy generalise instead of memorise.
Sourced collection
A collection engagement from your side of the table looks like this.
- 01
You define the brief
The task, the target skills, the conditions that matter and what counts as a good example. The brief is the contract for the run.
- 02
We select the stations
We match the brief to suitable stations across the network, usually across several plants, so the resulting data carries variety rather than one plant's habits.
- 03
Capture runs on your terms
On our capture kits, or on cameras and rigs you supply so the data arrives through the same optics your robot will use.
- 04
You receive structured datasets
Labelled, quality-scored and delivered in your format, with provenance and consent documentation attached.
- 05
You iterate on the brief
Review what came back, change what you are asking for, and run again. Nothing about the arrangement has to be renegotiated to do that.
Structuring your footage
If you already hold operator video, you do not need a collection programme. Send what you have and receive structured datasets back.
No new hardware
Works with the footage and cameras you already have. Nothing gets installed on a line for this.
No workflow change
Your teams keep working the way they work. The footage moves, the operation does not.
The same outputs
Identical fields and identical delivery formats to a sourced collection run, so the two can be mixed in one training set.
Quality handled for you
Unusable footage is filtered out and every delivered segment carries a quality score, so you are not training on the parts that should have been thrown away.
What you receive
The same set of structured outputs, whichever route the footage came in through.
- Temporal action segments
- Every discrete action in the footage, with precise start and end boundaries.
- Task and action labels
- A task label and an action type on every segment, so a segment is usable without watching the video.
- Object and hand masks
- Segmentation masks for the hands and for the objects being handled.
- Interaction and contact events
- Where and when contact is made with an object, and when it is released.
- Hand and wrist trajectories
- Continuous hand and wrist paths through each action.
- Spatial motion data
- Derived from standard 2D video. No depth sensors and no motion capture suits are required on the line.
- Per-segment quality score
- A quality score on every segment, with unusable footage filtered out automatically before delivery.
- Privacy flags and redaction
- Sensitive segments are flagged and redaction is applied before a dataset leaves us.
Delivery formats
Datasets are delivered in HDF5, LeRobot or RLDS. If your team works to its own schema, we deliver to that instead.
Quality and privacy handling
Quality scoring on every segment
Each delivered segment carries a quality score, so you can filter to the threshold your training run needs.
Unusable footage removed
Occluded, truncated and unrecoverable material is filtered out automatically rather than shipped and left for you to find.
Privacy flags travel with the data
Segments needing downstream care are flagged, so your own review can see what was treated and why.
Redaction before delivery
Faces and identifying features are handled before a dataset leaves us, not after it lands.
Why factories take part
Participating plants are not doing anyone a favour. Every one of them receives a process improvement report and an ergonomic analysis report at no cost, produced from the same operator footage, plus access to our commercially deployed process analysis platform. Operators are compensated for their participation, and capture is designed not to disturb the line.
More on our process intelligence platform for manufacturers(opens in a new tab)Scope a pilot on a sample
Send us a sample of footage, or a task you want captured. We will come back with what we can deliver, in what format, and what the consent position looks like.
Talk to our team