Evidence-Grounded Multimodal Semantic Driving (EGMS-Drive): Controlled Synthetic Mechanism Evaluation and Power-Informed Safety Validation Planning for Mixed-Traffic Driving

Abstract

: Mixed-traffic driving requires coherent perception and decision-making under sensor degradation, temporal variation, complex agent interactions, and multimodal futures. This paper presents Evidence-Grounded Multimodal Semantic Driving (EGMS-Drive), an explainable driving framework that integrates reliability-aware multimodal representations, cross-modal alignment, motion-compensated temporal consistency, temporal traffic scene graphs, probabilistic intent and trajectory inference, tail-risk-aware utility optimization, minimal sufficient evidence, and an evidence-constrained large language model. Safety-critical control remains numerical and nonblocking, while the language component verbalizes structured evidence produced by the multimodal perception and temporal scene-graph pipeline. The framework was examined through three controlled synthetic studies and a prospective power analysis. Study 1 reports a fully executable, post hoc exploratory comparison of Baseline B and a modified Structured fusion implementation across ten paired training-data and model-initialization replicates evaluated on fixed test and rollout tapes. Structured fusion increased macro-F1 by 0.0241 and reduced negative log-likelihood and the multiclass Brier score by 0.0383 and 0.0313, respectively; all three contrasts favored Structured fusion in 10/10 replicates and remained significant after Holm adjustment. Calibration and controlled synthetic rollout superiority were not established. In Studies 2 and 3, five experimental conditions were fitted independently across 10 training-sample replicates. In Study 2, reliability weighting reduced adverse-injection negative log-likelihood by 0.0080, cross-modal alignment increased the standardized alignment score by 2.6021, and temporal consistency reduced one-step feature-prediction error by 0.0548. In Study 3, the temporal-graph mechanism reduced the history-dependent minimum final displacement error over six candidate trajectories, minFDE 6 , by 0.3837 m relative to the no-graph condition and by 0.7285 m relative to the spatial-only graph, while predicted-intent gating reduced the six-mode Brier-adjusted minimum final displacement error, Brier-minFDE 6 , by 6.3895. A paired, cluster-aware power analysis further yielded a provisional requirement of 1260 episodes per condition for confirmatory collision evaluation. Overall, the results support improvements in action prediction, fusion reliability, cross-modal alignment, temporal representation, interaction modeling, and intent-conditioned multimodal trajectory forecasting and provide a statistically grounded design for subsequent confirmatory evaluation. However, they do not establish superiority in CARLA, real-road testing, or deployment safety.