Calibrated relational world model detects semantic mismatches in action interfaces

Rui Zhu, Xurui Li, and Haixu Tang propose an offline diagnostic method based on a calibrated relational world model that uses learned state-action-effect relations as a contract to detect robot action interface mismatches that are shape- and range-compliant but semantically wrong (axis permutation, sign inversion, rotational scaling, gripper polarity inversion), and to rank the fields likely to be at fault. On 10 LIBERO Spatial tasks and 432 source episodes, using capacity-controlled comparisons at a 2.5% empirical clean false-positive rate, the ridge-regression forward residual achieves 46.1% recall, the four-component joint Gaussian mixture 61.0%, and the exact conditional of the same mixture 65.7%. The authors state that these equal-rate comparisons are retrospective diagnostics and do not constitute a deployment calibration guarantee; DROID, three HYDRA tasks, short-duration faults, small offsets, exact repair, and simulated mitigation each failed their respective thresholds, and xArm contributed only one task. The authors therefore limit the conclusions to bounded offline diagnostics and a detection-localization trade-off, rather than joint-density superiority or cross-robot safety.

AI 根据报道生成 · 6 days ago更新