Stop copying fingertip drafts: a contact-aware controller lifts real-world teleoperation success from 29.2% to 78.3%

Dexterous hands can receive useful motion intent from teleoperation or learned policies, yet whether a motion succeeds also depends on contact decisions those commands rarely specify. Such upstream commands have typically been treated as trajectories to copy, leaving contact-level trade-offs to the execution side. DexFLEX treats a draft as evidence about intent, and the same trained controller is compared across degraded-command simulation, real-world shared control, and visuomotor policy learning.

From the current tactile-proprioceptive state it proposes several feasible short-horizon motion chunks, predicts their contact consequences, and selects the one that best follows the command while preserving future contact stability. At inference a Draft-Dream-Select procedure combines pure-prior proposals for recovery with draft-seeded proposals for responsiveness. The selected motion is decoded into executable joint-space targets; real-world teleoperation success rises from 29.2% to 78.3%, and policy-learning success reaches 46.7%, 3.5 times direct joint-action prediction and 1.56 times prior-only correction.

This report covers the summary only: DexFLEX turns upstream fingertip-motion drafts into contact-consistent joint commands, validated in degraded-command simulation, real-world shared control, and visuomotor policy learning. It suits readers working on dexterous manipulation, teleoperation, and policy learning, and engineers who need upstream motion intent to reach the execution side. The summary states no next step.

The summary gives success rates but not how commands were degraded, the real-world tasks or hardware, or sample sizes per setting; those details need the full text.