Event

PRIME learns precise progress estimation from imprecise demonstrations

1 reports · 1 sources

Huanqian Wang et al. propose the progress estimator PRIME, used to generate dense reward signals in robot manipulation reinforcement learning, learning precise task progress estimation from imprecise demonstration data. PRIME builds a weighted graph over observation states, connecting states across trajectories by spatial proximity and visual consistency, with edge weights encoding spatial distance as transition cost, and uses shortest-path search to approximate the geodesic distance to success; its variation can distinguish progress, stagnation, and regression, without assuming progress is monotonic over time. The paper states that in experiments on precision manipulation tasks in simulation and the real world, PRIME performs consistently effectively across a variety of tasks and is robust to data of different degrees of suboptimality; on the charger insertion task, its policy success rate improves by up to 5 times compared with the SFT baseline, and the advantage is more pronounced in higher-precision scenarios.