From hand-coded geometry to learned policies
Early grasp planning systems relied on hand-coded geometric rules: measure the object's edges, compute force closure analytically, and select contact points using explicit physics equations. This works reasonably well for objects with known, simple geometry, but it breaks down fast for the messy variety of shapes a real warehouse or kitchen actually contains. Writing geometric rules to cover every possible irregular shape isn't just tedious, it's fundamentally limited, because no fixed rule set anticipates every object a robot will eventually encounter.
Modern grasping systems increasingly use models trained on large datasets of grasp attempts, both successful and failed, often gathered from thousands or millions of simulated or real trials. Instead of computing force closure from an explicit geometric model, the system learns statistical patterns about what contact points and approach angles tend to succeed given an image of an object, and it outputs grasp candidates directly from that learned experience. This is conceptually similar to how general-purpose physical AI models learn broad policies from large amounts of interaction data rather than being explicitly programmed for every scenario.