Lazy-MOSAIC: Accelerating Skill-Centric Manipulation Planning with Learned Skill-Effect World Models
Every skill in a returned plan is checked in simulation; learned predictions help decide which candidates are worth checking.
Lazy-MOSAIC predicts a robot skill’s feasibility and resulting world state in a fraction of a millisecond. It uses those predictions to focus expensive simulation on promising candidates, repairing only the affected search region when a prediction fails. On cluttered-shelf retrieval tasks, the authors report substantially faster planning and fewer simulator queries than baselines, with equal or better success rates in every tested setting. The paper also includes a physical-robot demonstration and is under review.
ArXiv · AI/CL/LG's note
Lazy-MOSAIC predicts a robot skill’s feasibility and resulting world state in a fraction of a millisecond. It uses those predictions to focus expensive simulation on promising candidates, repairing only the affected search region when a prediction fails. On cluttered-shelf retrieval tasks, the authors report substantially faster planning and fewer simulator queries than baselines, with equal or better success rates in every tested setting. The paper also includes a physical-robot demonstration and is under review.
ArXiv · AI/CL/LG's note
score 4