This report presents our second‑place solution to the HANDS 2026 workshop challenge (Dexterous Grasp Motion track). The task requires generating physically plausible reach‑and‑lift trajectories for the 12‑DoF LinkerHand O6 on unseen objects in simulation, starting from random initial hand poses. Each time step involves roughly $70$ twelve‑dimensional decisions, so errors accumulate, and test objects and dynamics may differ from training, making the problem especially hard. Instead of generating motions step‑by‑step, we edit a single successful GraspM3 demonstration: the policy observes the target object once, outputs a 12‑D warp vector that deforms the demonstration, and then replays the warped trajectory open‑loop. The warp policy is trained with one‑step PPO over all $4{,}824$ training objects in parallel. Our method achieved a $94.61\%$ success rate on the easy track—the highest among submissions—and $57.18\%$ on the hard track of the private test set.
Review: The single‑shot warping approach effectively mitigates error accumulation, and parallel PPO training enhances generalization, offering a strong solution for high‑dimensional dexterous grasp generation.