FlashDexRetarget

FlashDexRetarget is a research method for physics-based retargeting of human hand–object demonstrations to dexterous robot hands. Instead of optimizing each demonstration separately, it trains a single reference-conditioned reinforcement learning policy jointly across a whole collection of motions, and stores every successful simulated rollout as robot data. The aim is to cut the cost of generating dexterous manipulation data.

The policy observes the simulated state, object point clouds, signed-distance-based hand–object features, and a latent encoding of the next 10 reference frames. Rewards use object point-cloud matching and a reference distance term for fingertips. Each hand has its own actor and critic, and training uses the off-policy FlashSAC algorithm scaled to the multi-reference setting.

On a 50-motion benchmark (HOT3D, TACO, OakInk2) retargeted to XHand and Sharpa Wave Hand in IsaacSim, the project page reports about 90% success on XHand with roughly 100x less training compute than CHORD, and compares against Do as I Do, DexMachina and CHORD. It also reports scaling to collections of 200, 500 and 1,000 motions, and a real-world replay of retargeted trajectories on a robot arm with an XHand. A video summary says code has been released and training needs about 110 GB of GPU memory; the code link was not in the provided material.