Multi-Skill Dexterous Manipulation
From RL to flow-matching — a three-stage pipeline for dexterous grasping
Dexterous grasping fails in a characteristic way: a policy discovers one grasp that works and collapses onto it, losing the variety that makes a hand worth having. This project builds a three-stage pipeline designed around that failure.
Advised by Dr. He Wang, Jan 2026 – May 2026.
The pipeline
1 — Flow-matching pose proposer. A flow-matching model conditioned on the object point cloud proposes grasp poses. Because it models the full distribution rather than a point estimate, it keeps multiple valid strategies alive for the same object.
2 — RL pick-up policy. An RL policy learns to execute a pick-up, rewarded against the proposed grasps. Grounding the reward in the proposer’s distribution is what prevents mode collapse — the policy is paid for realizing the variety the proposer offers, not for finding one grasp that always works.
3 — Distilled visuomotor student. A visuomotor policy is distilled from the RL teacher’s demonstrations, so the final policy runs from vision alone.
Fixing behavioral cloning’s blind spot
A distilled student only ever sees states the teacher visits, so the first small deviation puts it somewhere it has no training signal for. I injected noise during demonstration collection to push the teacher off-distribution and record how it recovers, giving the student the corrective behaviors that clean demonstrations never contain.
The pipeline reached 85%+ success across Dexonomy, retaining diverse grasp strategies rather than converging on one.
