Robots struggle to learn dexterous manipulation because the contact-rich experience required for such skills is hard to collect autonomously. In a new arXiv preprint (2610.22340), researchers propose DSP, or Dexterity from Supervised Play, a framework in which a robot first learns a play policy from minutes of human play. That policy is then used to generate the contact-rich experience needed for downstream manipulation tasks.
By framing data collection as play rather than task-specific demonstration, DSP aims to avoid costly manual engineering and expert teleoperation. The approach is called supervised because the play data comes from human operators, but the resulting policy is intended to produce diverse interactions on its own.
The abstract does not include quantitative results or benchmark comparisons, so the claims are limited to the proposed framework itself.