Dexterous-hand retargeting drops per-object training: one residual policy grasps unseen objects zero-shot

Physics-based retargeting for dexterous hands can now be done by a single unified policy instead of one trained per object. Earlier approaches either omitted physical constraints or relied on computationally expensive sampling-based optimization in simulation, while standard reinforcement-learning policies required object-specific training and failed to generalize across arbitrary geometries. Rolling the policy out in simulation along a kinematic reference trajectory yields a retargeter that recovers 58% of held-out and 54% of unseen-object DexYCB trajectories while keeping tracking error under 2 cm.

Zero-shot generalization across objects depends on the structure of the observation space rather than on policy scale. The observation space is formulated to include a localized contact-error signal that guides each robot link toward its target coordinate on the object surface. This residual policy runs in simulation at 0.4 seconds of compute per second of reference, reaching 2.4 times faster than real time and a 60 to 350 times reduction in computational cost relative to sampling-based methods.

This report covers what the summary states: the system rolls the policy out in simulation along a kinematic reference trajectory to produce the physics-based retargeter OATH, evaluated on DexYCB trajectories. It is aimed at researchers and engineers working on dexterous manipulation, motion retargeting, and physics-based animation. The summary states no next experimental step.

The reported results are from simulation: recovery is 58% on held-out and 54% on unseen-object trajectories, with tracking error held under 2 cm, and no real-hardware evaluation is described. The summary does not say under which object categories or contact conditions these fractions would drop.