Hao Shi

Driver workload classification requires a connection between observed multimodal responses and the cognitive demands of driving tasks. We propose ACT-R cognitive semantic lightweight fusion (ACTR-CSLF), which summarizes cognitive model states and production rules into 20 interpretable task semantic features and integrates them with 16 driving, eye-tracking, and electroencephalographic features. Evaluation covered 420 segments from 30 participants across seven simulated highway tasks and two weather conditions. In leave-one-subject-out validation, Human+ACT-R achieved Macro-F1 = 0.7277, compared with 0.5795 for Human-only and 0.7485 for ACT-R-only. Class-wise results showed improvements over Human-only at all three workload levels. Modality ablations identified input-dependent complementarity, while condition-level associations connected task semantics with subjective workload and physiological responses. Under simultaneous participant-and-scene holdout, pooled Macro-F1 was 0.5439 for Human+ACT-R, 0.4757 for ACT-R-only, and 0.4092 for Human-only; scene-averaged results characterized the uneven distribution of these gains. Condition baselines and random code controls separated task prior prediction from semantic interpretation. The joint classifier used 111 parameters, 1.0879 KB of storage, and 0.2147 ms mean single-thread CPU inference time. ACTR-CSLF provides a compact interface between explicit cognitive model knowledge and individual multimodal responses, combining workload discrimination with traceable task semantics.

Ethical decision-making and responsibility assessment in autonomous driving are often addressed separately, which limits the consistency between maneuver selection and post-event analysis. This study proposes a responsibility-aware decision-making framework that integrates a multi-actor responsibility model with a comprehensive ethical-cost function. Moral-risk scenarios are classified and represented using variables related to the vehicle state, road environment, traffic participants, human–machine interactions, and data traceability. Seven responsibility factors are weighted using expert elicitation and the analytic hierarchy process. Responsibility shares for five actor groups are then calculated from the weighted contribution matrix. Candidate maneuvers are first screened by vehicle-dynamics constraints and then evaluated using five normalized cost components: collision risk, harm to life, rule violation, vehicle-stability risk, and responsibility transfer. The framework is evaluated in a pedestrian red-light crossing scenario through SCANeR–MATLAB co-simulation. Under the baseline condition, emergency braking produced the lowest comprehensive ethical cost of 0.276, while the external traffic participant and autonomous-driving system received relative responsibility weights of 0.50 and 0.26, respectively. Across five operating conditions, the preferred maneuver varied between emergency braking and deceleration while maintaining the lane as the vehicle speed and surrounding traffic gaps changed. The results indicate that the framework can adapt maneuver selection to scenario-dependent risk while maintaining a traceable assessment of relative responsibility contributions. The framework provides a quantitative engineering approach for integrating ethical maneuver selection with responsibility assessment; the resulting responsibility weights are scenario-dependent analytical quantities and are not intended to determine legal liability. These results may inform the design and simulation-based evaluation of responsibility-aware autonomous-driving strategies.

Navigate-on-Pilot (NOP) systems can experience delayed awareness of interacting vehicles in traffic conflict scenarios because of the limited effective perception range of onboard sensors. To improve the safety of NOP longitudinal control, this study develops a safety-driven cruise speed control method for V2X-enabled NOP vehicles. Systems-Theoretic Process Analysis (STPA) is used to identify unsafe control actions, their causes, and corresponding safety goals. V2X information is incorporated to extend vehicle perception, while communication delay is considered when updating interacting-vehicle states. A two-dimensional safety-envelope model is then used to detect potential spatial conflicts, and the remaining path distance to the predicted conflict region is compared with the minimum stopping distance to determine emergency braking. Real-vehicle experiments were conducted in T-intersection and roundabout scenarios under potential-conflict and non-conflict conditions. The proposed method distinguished interactions with different spatial and temporal relationships under both straight and curved paths. Compared with an onboard-perception-based strategy, the V2X-enabled strategy advanced the braking response by approximately 1.6 s at the T-intersection and 2.4 s in the roundabout. The results indicate that earlier availability of interacting-vehicle information can increase the time available for conflict assessment and longitudinal control without requiring substantially greater braking intensity.

To address the challenges of recognition lag and limited representational capacity of single-modal physiological information in driver intention recognition under human–machine co-driving environments, this study constructed a highway driving takeover experimental scenario. Typical takeover tasks including obstacle avoidance takeover, adjacent vehicle cut-in takeover, and curved section takeover were designed, with simultaneous collection of drivers’ electroencephalography (EEG) and eye-tracking data. On this basis, a multi-modal physiological information fusion method for driving intention recognition was proposed. First, areas of interest (AOIs) were defined to extract visual behavioral features, including fixation duration, fixation count, and pupil diameter, while eye-tracking recurrence plots were generated based on phase space reconstruction. Meanwhile, event-related potential (ERP) analysis was performed on the EEG data to extract N100, N200, and P300 components, and continuous wavelet transform was applied to construct EEG time–frequency maps. Subsequently, the EEG time–frequency maps and eye-tracking recurrence plots served as the dual inputs to the dual-branch Squeeze-and-Excitation Convolutional Neural Network (SE-CNN) model, which was further enhanced by fusing ERP features and eye-tracking statistical features, ultimately achieving accurate recognition of four types of driving intentions. The model results demonstrated that the dual-branch SE-CNN model achieved a recognition accuracy of 88.29% on the four-class intention recognition task, with precision, recall, and F1-score reaching 88.78%, 88.26%, and 88.31%, respectively—significantly outperforming both single-modal models and the dual-branch Convolutional Neural Network (CNN) model without the SE attention mechanism. The proposed model can effectively integrate EEG neural activity signals and eye-tracking visual behavioral information, thereby substantially enhancing driving intention recognition performance. This provides theoretical foundations and technical support for applications such as early driver intention perception, intelligent takeover decision-making, and human–machine collaborative control in human–machine co-driving environments.

At autonomous intersections in mixed traffic, where Connected and Autonomous Vehicles (CAVs) coexist with Human-Driven Vehicles (HDVs), scheduling must remain effective even when a reliable explicit HDV turning intention is not available sufficiently early for the scheduling decision. This paper proposes a learning-based platoon scheduling method for this information-limited setting. CAVs and HDVs are organized into mixed platoons or an HDV group, and a state-augmentation function is designed to encode their temporal relationship while preserving the priority of uncontrollable HDVs. An enumeration (EN)-based expert demonstration mechanism is further integrated into the training process to provide high-quality experience. In the representative training run reported in the manuscript, the expert-assisted model achieved a peak average reward approximately 14% higher than the model trained without demonstrations. Across the tested demand cases, the learned scheduler also obtained travel-cost performance close to the EN benchmark. The scope of the method is explicitly limited to the modeled lane-keeping and sensing assumptions; robustness to random seeds, unexpected HDV maneuvers, communication delays, and additional safety metrics requires dedicated validation.

Conventional multimodal driver-workload recognition relies on behavioral and physiological statistics, whereas discrete task labels do not explicitly encode risk, stage, or strategy. We propose ACT-R Cognitive Semantic Lightweight Fusion (ACTR-CSLF), which aggregates dynamic states from an ACT-R driving model into 20 task-cognitive semantic features and combines them with 16 driving, eye-tracking, and electroencephalographic features in multinomial logistic regression. Leave-one-subject-out validation used 420 weather-task segments from 30 participants, with condition-prior, one-hot, random-code, rank-matched, shuffled-semantic, and leave-one-scene-out controls. Macro-F1 was 0.5795 for Human-only, 0.7485 for ACT-R-only, and 0.7277 for Human+ACT-R; Condition-only reached 0.7674, revealing a strong closed-set task prior. Standard random codes averaged 0.7309, so superiority to arbitrary high-dimensional task codes was unsupported. However, the real ACT-R representation exceeded the empirical 95% upper bound of rank-matched random codes (0.7277 vs. 0.7251; one-sided empirical p = 0.0198), and all 30 shuffled mappings performed worse. In leave-one-scene-out validation, Human+ACT-R outperformed discrete unknown fallbacks in Macro-F1 (0.2590 vs. 0.1902-0.1929) and accuracy (0.5643 vs. 0.2690-0.2714). With 111 parameters, 1.0879 KB storage, and 0.2147 ms mean CPU latency, ACTR-CSLF offers a lightweight, traceable balance of predictive performance, cognitive interpretability, and deployment cost.