Planner-verified counterfactual tests reveal grasp predictor flaws

Sajid Ansari, Antony Thomas, and Girish Varma propose a planner-verified counterfactual (CF) test to audit grasp feasibility predictors: fixing the grasp pose, inserting physical obstacles in the approach corridor, and adding obstacle points to the model input, with the planner re-determining the label so that it flips from feasible to infeasible. Across eight predictors covering point clouds, voxels, depth, and object tokens, the CF paired AUROC drops from 0.920–0.996 in natural scenes to near-random 0.470–0.504. The authors state that this flip holds across four planners, two robotic arms, and real object obstacles; these predictors were trained on unmodified benchmarks, had not seen counterfactual scenes, and the tests remain limited by the obstacle placements used in their training.

AI 根据报道生成 · 2 days ago更新