Obstructed-Premise Commitment Failure in AI-Guided Autonomous Navigation: An FCL-S/PIB Framework for Non-Commitment under Adversarial AI Interference

Engineering notes

Key topics: autonomous driving, carla, perception, planning, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This paper introduces Obstructed-Premise Commitment Failure (OPCF) as a structural failure mode in AI-guided autonomous driving and autonomous navigation systems. OPCF occurs when an autonomous navigation system converts reasoning based on an adversarially obstructed, unverifiable, source-contaminated, conflicting, or expired premise into real-world navigation action without effective premise re-validation at the action boundary. The paper extends Hiroko Konishi’s framework of Premise Integrity Blindness (PIB), False-Correction Loop (FCL), and False-Correction Loop Stabilizer (FCL-S) to autonomous navigation. Its central claim is that safety in adversarially obstructed autonomous navigation should not be evaluated only by whether a system reaches its destination despite interference. It should also be evaluated by whether the system can refuse destination-seeking commitment when the premises supporting that commitment are obstructed, unverifiable, or provenance-corrupted. The paper proposes an FCL-S Navigation Governance Layer (FCL-S-NG). This layer represents sensor inputs, map updates, V2X messages, external AI commands, memory-derived assumptions, and mission goals as premise slots. Each premise slot preserves semantic content, source, freshness, provenance state, validation state, and operational criticality. Before a candidate navigation action crosses the commitment boundary, FCL-S-NG applies a Premise Integrity Gate, a Stop/Correction Boundary for Navigation (SCB-N), an Unknown Stable Terminal for Navigation (UST-N), and Attribution Fixation for Navigation (AF-N). The paper positions OPCF relative to adjacent work on runtime assurance, SOTIF, RSS, V2X trust management, adversarial perception, uncertainty-aware planning, and noncommitment. While these areas address safety envelopes, fallback control, formal driving rules, source trust, attack detection, and delayed planning decisions, OPCF formalizes a distinct commitment-safety problem: whether a represented premise is entitled to authorize physical action at all. The paper further proposes measurable evaluation metrics, including Ungrounded Navigation Commitment Rate (UNCR), Premise Re-Validation Rate (PRR), Safe Non-Commitment Rate (SNCR), False Continuation Rate (FCR), and Provenance Loss Rate (PLR). These metrics can be instrumented in CARLA-class simulations or autonomous-driving stacks by logging premise traces, dependency sets, validation states, and commitment-boundary events. This work does not propose tactical evasion methods against military AI systems or offensive countermeasures. Its contribution is a safety-governance formalization for preventing autonomous systems from operationalizing contaminated or unverifiable premises under adversarial AI interference.

Links and sources

Comments