Researchers led by Jinzhou Li released ReDex, a method for adapting dexterous hand policies trained in simulation to real hardware. The system lets a human gently correct finger contacts while the original policy runs the other fingers. Rollouts that mix the base policy, human‑guided finger moves, and fingertip force data are used to train a new force‑conditioned policy via behavior cloning. In tests on two contact‑rich tasks, ReDex raised object‑flipping success from 14% to 86% and screwdriver rotation progress from 26% to 95.3%.
Why it matters
It shows a practical way to turn simulation‑trained hand skills into reliable real‑world robot actions with minimal human correction.
