Reliability-Gated Multimodal Generative World Modeling with Conformal Distributional Reinforcement Learning for Whole-EV Optimization under Unseen Driving Scenarios

Abstract

The use of predictive models for the simultaneous representation of vehicle energy consumption, thermal load, mechanical and chemical ageing, traction, route, and passenger comfort is an essential condition to optimize electric vehicles operating in unpredictable environments. However, most existing studies are based on deterministic predictions, on predefined scenarios, on independent control of each subsystem, and on reinforcement learning with expected return, thus limiting their robustness when faced with multiple changes in the underlying probability distributions. Therefore, this work presents a sequential multi-modal generative approach combining five different analytical methodologies. Firstly, the Reliability-Aware Fusion (RGM-SCFT) method allows the integration of multirate information coming from various sensors (battery, motor, tires, camera, GPS, weather, driver). Secondly, the Counterfactual Action Generative model with Deterministic Wind Model (CAG-DWM) provides physics-consistent action-conditioned counterfactual futures, separating epistemic from aleatoric uncertainties. Thirdly, the Outlier-Rare Event Generation Curriculum (OREG-Curriculum) creates real rare event and out-of-distribution scenarios. Fourthly, the Conditional Quantile Regression Deep Reinforcement Learning (CQR-DRL) method learns optimal decision-making strategies by optimizing multivariate return distributions using conformal quantiles and conditional Value-at-Risk. Lastly, the Digital Twin-Test Hardware In-The-Loop Intervention Validation (DT-THIV) ensures policy validation using digital twin simulations, hardware In-the-loop tests, and intervention-based Validation in practical scenarios. The proposed sequential framework aims at providing reductions in traction energy consumption ranging from 9% to 14%, in degradation of up to 18–27%, in tire slip risk of up to 35–55%, in constraints violations of less than 3%, as well as a high level of simulation-to-reality transferability ranging from 90% to 95% while maintaining travel time and safety calibration for previously unseen environmental conditions.