From False Positives to Failure Topologies: Heuristic Learning for Auditable Dexterous Grasping Development

Early-stage dexterous grasping on a custom robot hand can fail before neural policy learning begins:the simulator, action interface, contact evidence, logs, and success criteria are still being constructed.We study Heuristic Learning (HL) as an auditable policy-debugging layer in which a coding agent editsa software policy and experiment harness after reading MuJoCo logs, failure reports, replay traces,and regression checks. A strict V3 rebuild reclassified selected demonstrations and assisted/stabilizedlong holds as false positives under pure-physics review: a controlled 450-row baseline across five policyfamilies produced 0/450 strict 15-second grasp successes, and the final HL policy produced 0/150 strictsuccesses in a 50-seed-per-object high-N evaluation. We use this strict boundary to define a V4 modularcontact-formation pilot: under a locked no-stabilizer evaluator, the V3 final-HL baseline reaches 0/50 cubecontact successes, a naive contact-only baseline also reaches 0/50, and a topology-guided V4 policy reaches50/50 on the same held-out seeds. This is a single-object-family, contact-only pilot; no long-hold, broadobject-generalization, or sim-to-real claim is made. The result shows that HL can expose false positives,build strict evaluators, localize failure modes, and convert failure topology into a targeted modularcurriculum. We position HL as a pre-neural debugging and curriculum-construction layer for customdexterous-hand development, complementary to reinforcement learning, imitation learning, diffusionpolicies, and large robot foundation models. More broadly, the study is a robotics instance of specificationauditing: a metric that once looked solved is shown to be an assisted proxy, then repaired into a stricterevaluator and a narrower module gate.