Mind the Control Gap: Robot Learning from Mismatched Data

Kushal Kedia

Keywords: RL, Sim2Real, Adaptation

TL;DR: Learning dexterous manipulation from mismatched data sources by treating the control gap as a design constraint

Abstract:

Modern robot learning systems map data to control, but the choice of data source strongly shapes what robots can learn. Teleoperation, the dominant paradigm, struggles with contact-rich tasks requiring forceful interaction and rapid contact switches. Human videos and simulation can more naturally capture such behaviors, but both are mismatched control data: they demonstrate what should be achieved, but not in the same control space or execution regime as a real robot.

In this talk, I will argue that this control gap should be treated as a first-class design constraint. First, I will present SimToolReal shows that a sufficiently general simulation setup can induce real-world dexterous tool-use without task-specific engineering to construct reward functions and environments. In the end, I will finish with a vision of lifelong adaptation, where robots continuously refine behavior after deployment to progressively close the mismatched control gap.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 4