GBPP: Two-Stage Learning for Robot Grasp Base Placement

A team of researchers has introduced GBPP (Grasp-Aware Base Placement Prediction), an efficient learning-based scorer that determines an optimal robot base pose for grasping based on a single RGB-D image. The approach consists of a two-phase curriculum: initially, a basic distance-visibility rule automatically labels a vast dataset cost-effectively; subsequently, a smaller collection of high-fidelity simulation trials fine-tunes the model to align with actual grasp results. Utilizing a PointNet++ style point cloud encoder alongside an MLP, the system evaluates dense grids of potential poses, allowing for quick online selection without extensive task-and-motion optimization. In both simulated environments and with a real mobile manipulator, GBPP surpasses proximity and geometry-only benchmarks, selecting safer, more accessible positions and demonstrating graceful degradation when errors occur. The findings provide an effective strategy for data-efficient, geometry-aware base placement: leverage low-cost heuristics for coverage, followed by calibration through focused simulation.