DexSeed expands 19 human grasp sequences into robot training data

DexSeed turns 19 human interaction sequences into 25,270 robot trajectories. Its researchers demonstrate a robot performing language-specified grasps after training on generated demonstrations.

Collecting examples often involves repeatedly teleoperating a robot. The November 10 preprint instead expands human interactions in physics simulation to produce training data.

This is not general household competence learned by watching anyone. The work centers on grasping and placement; its project page also shows a mug handle slipping and placement failing.

Reusable data could reduce repeated collection on real hardware. Reaching dependable household assistance would still require broader tasks and real-world testing.

Source published: arXiv, November 10, 2026, 21:52:18 Beijing / 08:52:18 EST. The team project page has no publication timestamp.