Collision-aware data structures to prune globally colliding grasps during search

Develop collision-aware data structures for the contact-field–based Lightning Grasp search pipeline that can proactively detect and prune candidate object placements and contact points which would lead to global hand–object penetrations, particularly for highly non-convex objects, during early search stages to maintain high effective sampling throughput.

Background

The paper observes that effective samples per second (SPS) drop markedly for highly non-convex objects (e.g., cups). Even though kinematics optimization resolves local collisions around each contact point under local convexity assumptions, global-scale hand–object penetrations still occur, producing many rejected samples.

The authors hypothesize adding finger-shape information into contact-field boxes to filter contact points that would cause collisions and note that earlier search stages should be collision-aware. They explicitly state that designing data structures to prune these failure cases during search remains open, motivating a targeted research direction to preserve throughput on challenging geometries.