Capture the professional skills your robot needs.
How experts perform the work, make decisions and judge the result, turned into tasks, training data and evaluations in simulation environments we build for the task.
Four services, scoped around one task.
You keep your model and training pipeline.
Your team shares the skill you need and a checkpoint if you have one. We build the environment and return recordings, annotations and evaluation results.
Start with one task and a delivery you can evaluate.
- Step 1
Scope the task
Bring the skill, and failure examples if you have a policy. We scope the environment and the first deliverable.
Task spec- SkillSeat a connector
- Failure examplesOptional
- First deliverableDemonstrations
- Step 2
Set up a sample session
Confirm robot controls, observations and task checks. Inspect a sample recording.
✓ Controls✓ Observations - Step 3
Run and collect
Experts record demonstrations. If you share a checkpoint, we also run it and experts correct its failures.
Collected - Step 4
Train and evaluate
Your team trains its model and compares the result on the reserved scenarios.
Reserved scenarios- Current policy
- After training
- Step 5
Decide what comes next
Extend the work once the pilot shows a useful result and a repeatable process.
Pilot review✓ RepeatableExtend the work
Questions
labs ask
Something else? Write to [email protected].
Who runs the policy?
If you have one, we do, in simulation. You share a checkpoint, we run it across task variations, and experts correct its failures. You do not need a policy to start: new skills begin with demonstrations. We never train your model.
What kinds of tasks work best?
Physical skills where judgment matters: assembly, handling, inspection, lab work, food prep and recovery from things going wrong. If a skilled person would say "it depends", it is a good fit.
Can you use our existing simulation?
Yes. We usually build the environment for you, but we can also work in Isaac Lab, MuJoCo, ManiSkill, Genesis or your own setup.
How are experts vetted?
On the work they actually do: credentials, years in the trade and practical screening on tasks in their domain. Every contribution is reviewed before it ships.
How do we start?
Bring one skill your robot cannot do yet, or one it does unreliably. We scope a pilot around it together and agree on the evals that define success.
Discuss your first task.
Bring a skill you are developing or a failure you want to understand. We scope the first pilot together.