Correct your policy from the exact state where it failed.
Your policy runs in simulation. When it fails, an expert takes over from that state, recovers it and notes why.
PolicyGoes to insert with the strands still splayed.
Corrections where your policy actually fails.
You need a checkpoint that runs in the task's simulation. No policy yet? Start with demonstrations.
Intervention spans
The policy's attempt, the handover and the expert's recovery, marked in one episode.
Why it failed
What went wrong and what the right response is, in the expert's words.
- LeRobot
- RLDS
- HDF5 or yours
Ready to aggregate
Add it to your dataset and retrain. The pipeline stays yours.
The data DAgger asks for, from people who know the work.
A policy trained only on demonstrations drifts into states the demonstrations never covered, and small mistakes compound. DAgger (Ross et al., 2011) fixes this by running the policy, having an expert give the right action in the states it actually reaches, and retraining on the combined data. HG-DAgger (Kelly et al., 2019) makes it work with people: the expert only takes control when the policy goes wrong.
That is the loop we run for you. In simulation we can stop at the exact failure state and hand it to an expert, so every correction starts where your policy went wrong. Then measure the next version with an evaluation on scenarios it never trained on.
Run it, stop at the failure, hand it to an expert.
Repeat with each new checkpoint. Only reviewed episodes are delivered.
- Share
Your checkpoint, our simulation
We run your policy in the simulation we build for the task.
Task creationCheckpoint- Runs in the task sim
- Same cameras and state
- Situations agreed
- Run
Attempts across the agreed situations
Failures are selected for correction.
Scenarios running - Take over
From the exact failure state
An expert takes over where the attempt went wrong and recovers it.
Episode 0142Robot attemptExpert intervention - Annotate
Say why, as it happens
The expert notes what went wrong and the right response.
Expert notesSplayed strands, re-twist - Deliver
Reviewed intervention spans
As LeRobot, RLDS, HDF5 or your format, with usage rights agreed upfront.
Episode review✓ Included
Questions about corrections.
Something else? Write to [email protected].
What is DAgger in imitation learning?
DAgger (Dataset Aggregation) is a way to fix a policy trained on demonstrations. You run the policy, have an expert give the right action in the states it actually reaches, add those labels to the dataset and retrain. HG-DAgger is the human-gated version: the expert only takes control when the policy goes wrong.
What is intervention data?
Robot episodes where a person takes over from the policy partway through. Each one records the policy's attempt, the handover and the person's recovery, so you can train on corrective actions in the states your policy actually gets into.
Do we need a policy to start?
Yes. Corrections need a checkpoint that runs in the simulation for the task. Without one, start with expert demonstrations and add corrections once you have a policy.
Who makes the corrections?
Skilled professionals from the trade the task comes from, such as electricians, mechanics, machinists and sonographers. Every expert learns the teleoperation controls on practice tasks first, and every episode is reviewed before delivery.
Do you train the model?
No. You keep your model and training pipeline. We return the intervention data; you aggregate it with your existing data and retrain.
Which formats and simulators do you support?
Data is delivered as LeRobot, RLDS, HDF5 or your own format, with usage rights agreed upfront. Tasks run in Isaac Lab, MuJoCo, ManiSkill, Genesis or your own simulator.
How do we know the corrections helped?
Run each policy version on a reserved test set that never appears in training, scored against expert-defined criteria. That is our evaluation service, and it can be scoped alongside corrections.
Bring a checkpoint and a failure.
We run it in simulation and return reviewed intervention spans with the expert's reasoning, as LeRobot, RLDS, HDF5 or your format.