Retargeting Is the Strategic Layer of Wearable Capture
The mapping from human motion to robot action is not a lossy compromise. Done right, it is the layer that lets one dataset feed every embodiment a lab ever ships.
The most common objection to wearable capture is that human hands and robot hands are different, so retargeting introduces error. It is true that the two are different. It is also true that retargeting is a solved-enough problem for most target morphologies, and that framing it as pure loss misses the strategic value of having it in your pipeline in the first place.
The framing that gets it wrong
'We can't use wearable data because our robot hand doesn't have five fingers.' This treats the human demonstration as raw material that has to fit the robot exactly, and it forgets that the same objection applies to teleop from a different arm, those demonstrations also do not fit exactly, and no one considers them useless.
The framing that gets it right
The human demonstration captures the intent and the fine structure of the skill. The retargeting layer maps that intent onto the target embodiment. If the target changes, the retargeting layer changes; the underlying demonstration does not. This is the same reason model teams write their training pipelines as configurable stacks rather than hard-coding for one architecture, the abstraction pays for itself the first time you swap a component.
For a lab that plans to ship more than one robot in its lifetime, this abstraction is the difference between a data asset that appreciates and one that depreciates. Wearable-captured hours accumulate; the retargeting layer evolves; the same session that trained your first policy can be re-retargeted to train the third one.
What good retargeting looks like
It exposes the raw signal cleanly, joint angles, contact events, force per fingertip, so any team can build their own map. It ships reference retargeters for the most common hand morphologies so teams that do not want to write their own can move fast. It is versioned, so retraining with an updated retargeter is reproducible. And it never hides the loss it introduces; every retargeter reports its error against ground truth so downstream teams know what they are working with.
This is the design brief for the BLO LAB retargeting stack, and it is why we treat it as a first-class product rather than a helper script.
Building or training robots?
We license manipulation datasets and run custom capture programs. Get in touch to see what fits.
