Skip to main content
Back

KAIST's FlashDexRetarget Teaches Robot Hands 90% Success in 29 GPU-Hours

A single reinforcement learning policy retargets hundreds of human hand demonstrations to dexterous robot hands, hitting 90% success with 100x less compute than prior methods.

Paper, Full Access
  • FlashDexRetarget trains one RL policy to retarget many human hand demos to dexterous robots at once.
  • Hits 90% success on 50-motion benchmark with XHand in 29 GPU-hours, versus CHORD's 46% at 2,847 GPU-hours.
  • Scales to 1,000 motions, converting 621 with one shared policy in 600M steps.
  • Uses point clouds, signed distance features, and future reference encoding as observations.
  • Separate left and right hand actor-critics plus scaled FlashSAC off-policy learning drive the gains.
  • Paper and project page live at davian-robotics.github.io/FlashDexRetarget, code coming soon.

Teaching robot hands to manipulate objects often begins with human demonstrations. Retargeting converts those recordings into robot trajectories that preserve the object motion and hand-object contacts while remaining physically valid in simulation. Differences in joint counts, finger lengths, and contact geometry make direct trajectory transfer unreliable, which limits the production of large dexterous-manipulation datasets.

Researchers from KAIST AI and Holiday Robotics have released FlashDexRetarget, a reinforcement-learning framework that trains across collections of demonstrations. On a 50-motion benchmark, the paper reports 90% success for XHand after roughly 29 GPU-hours. CHORD, the strongest prior reinforcement-learning baseline in the comparison, reached about 46% after 2,847 GPU-hours.

Reported results on the 50-motion XHand benchmark
Method Success Training cost
FlashDexRetarget 90% About 29 GPU-hours
CHORD About 46% 2,847 GPU-hours

One-motion training hits a wall

Physics-based systems such as Do as I Do, CHORD, and DexMachina optimize each demonstration independently. Their compute cost therefore grows with the number of source motions, while difficult clips can still fail after long training runs. That workflow becomes prohibitive when a large imitation-learning or manipulation model needs hundreds of thousands of physically valid trajectories.

FlashDexRetarget treats the task as multi-reference tracking. A reference-conditioned control system learns from many demonstrations during the same training run, spreading the optimization cost across the collection. Successful simulated rollouts become robot demonstrations that can be stored for downstream policy training.