Chen Sun

An autonomous vehicle can complete its route while disrupting surrounding traffic. We introduce SAFE-AD, augmenting reinforcement learning (RL) observations with compact descriptors of a finite-horizon interaction-pressure field constructed from current traffic states. The complete objective supplements the environment's task reward with risk, comfort, and externality costs. Then, we formulate a context-conditioned physics-informed neural network (PINN) to approximate the numerical field. Evaluation combines held-out naturalistic recordings, HighwayEnv diagnostics, and closed-loop MetaDrive tests in various traffic scenarios, with a matched factorial experiment with continuous soft actor-critic (SAC) separates observation and objective effects. At unprotected intersections, the complete profile reduces collision rate from 0.296 to 0.220 and mean absolute jerk by 0.98 m/s3 relative to baseline SAC. Besides, the mean absolute jerk in roundabout decreases by 1.54 m/s3. The results show scenario-specific gains and sensitivity to objective calibration.

This package contains the numerical propagated interaction-field model, context-conditioned physics-informed neural network (PINN), field-aware reinforcement-learning pipeline, and processed measurements supporting the principal validation and simulator findings. The statistical workflow preserves the original observation/reward factorial and replaces only its complete-model merge candidate with the selected 0.25-times comfort profile. All five calibrated merging Soft Actor-Critic checkpoints and their five paired stock checkpoints are included. Policy-training seeds are 0-4; each policy is evaluated on traffic seeds 10000-10049 at density 0.3. The calibrated profile uses jerk weight 0.000375 and steering/throttle-change weights 0.00025. An indexed mapping connects each checkpoint to its effective training settings and episode CSV. The PINN checkpoint is accompanied by its available architecture, input channels, normalization, numerical-solver settings, loss weights, calibration-cache signatures, hardware record, training duration and 6,000-step loss history.

Reinforcement learning (RL) has shown excellent performance in solving decision-making and control problems of autonomous driving (AD), which is increasingly applied in diverse driving scenarios. However, driving is a multiattribute problem, leading to challenges in achieving multiobjective compatibility for current RL methods, especially in both policy updating and policy execution. On the one hand, a single value evaluation network limits the policy updating in complex scenarios with coupled driving objectives. On the other hand, the common single-type action space structure limits driving flexibility or results in large behavior fluctuations during policy execution. To this end, we propose a multiobjective ensemble-critic (MoEC) RL method with a hybrid parametrized action for multiobjective compatible AD. Specifically, an advanced MORL architecture is constructed, in which the ensemble-critic focuses on different objectives through independent reward functions. The architecture integrates a hybrid parameterized action space structure, and the generated driving actions contain both abstract guidance that matches the hybrid road modality and concrete control commands. In addition, an uncertainty-based exploration mechanism that supports hybrid actions is developed to learn multiobjective compatible policies more quickly. The experimental results demonstrate that, in both simulator-based and HighD dataset-based multilane highway scenarios, our method efficiently learns multiobjective compatible AD with respect to efficiency, action consistency, and safety.

Sustaining long-term interactions remains a bottleneck for Large Language Models (LLMs), as their limited context windows struggle to manage dialogue histories that extend over time. Existing memory systems often treat interactions as disjointed snippets, failing to capture the underlying narrative coherence of the dialogue stream. We propose TraceMem, a cognitively-inspired framework that weaves structured, narrative memory schemata from user conversational traces through a three-stage pipeline: (1) Short-term Memory Processing, which employs a deductive topic segmentation approach to demarcate episode boundaries and extract semantic representation; (2) Synaptic Memory Consolidation, a process that summarizes episodes into episodic memories before distilling them alongside semantics into user-specific traces; and (3) Systems Memory Consolidation, which utilizes two-stage hierarchical clustering to organize these traces into coherent, time-evolving narrative threads under unifying themes. These threads are encapsulated into structured user memory cards, forming narrative memory schemata. For memory utilization, we provide an agentic search mechanism to enhance reasoning process. Evaluation on the LoCoMo benchmark shows that TraceMem achieves state-of-the-art performance with a brain-inspired architecture. Analysis shows that by constructing coherent narratives, it surpasses baselines in multi-hop and temporal reasoning, underscoring its essential role in deep narrative comprehension. Additionally, we provide an open discussion on memory systems, offering our perspectives and future outlook on the field. Our code implementation is available at: https://github.com/YimingShu-teay/TraceMem

Decision-making and control are the core functionalities of high-level autonomous driving systems. Existing mainstream research, including modular and end-to-end paradigms, typically employ conservative strategies that compromise driving efficiency. However, driving efficiency constitutes a critical constraint on the transition of autonomous vehicles from mere operability to practical utility. Autonomous overtaking systems serve as a typical means to improve driving efficiency. Nevertheless, in stochastic and uncertain traffic scenarios, achieving safe and efficient continuous autonomous overtaking remains a significant challenge. In this context, this paper proposes a decision-making and control framework based on MeUAL to achieve the optimal trade-off between overtaking risk and efficiency. First, at the decision-making layer, a safe reinforcement learning method based on Uncertainty-aware Augmented Lagrangian (UAL) is developed to provide global overtaking guidance. Subsequently, the motion planning and control layer based on Model Predictive Control (MPC) closely tracks the UAL-generated guidance, while preserving the safety and constraint guarantees inherent to traditional MPC. Finally, a Policy Switching Mechanism (PSM) triggered by the safety epistemic uncertainty threshold is designed for the MeUAL-driven autonomous overtaking system. Experimental results demonstrate that MeUAL outperforms baseline algorithms with respect to reward-cost balance, sample efficiency, and learning stability. Moreover, in various test scenarios that are distinct from the training distribution, MeUAL-PSM exhibits strong robustness and interpretable overtaking maneuvers through flexible policy switching.