Xiaocong Zhao

Surrounding vehicles may misjudge an autonomous vehicle’s intended motion and yield or merge at the wrong time. These conflicts are difficult to test systematically when surrounding vehicles’ expectations cannot be varied separately from the autonomous vehicle’s planned motion. This paper proposes an intention-misjudgment-oriented background-traffic generator for autonomous-vehicle testing. A speed-intention perturbation mechanism separates the nominal speed plan of the vehicle under test (VUT) from the estimate supplied to surrounding vehicles, enabling controlled underestimation and overestimation. A Transformer-based interaction encoder relates the perturbed estimate to traffic context and 50-step guiding-lane references to generate context-dependent responses. Multimodal decoding represents alternative responses, while learned affine state updates couple their evolution to inter-vehicle interactions and VUT control inputs. In highway-merging simulations, the best-matching generated trajectories have lower average and endpoint position errors than those of the Mixtures of Affine Time-varying Systems (MATS) baseline trajectory-prediction model at all three comparison speeds. Across 2353 matched initial scenarios, underestimating the VUT’s future speed by 25% yields an observed collision rate of 11.3%, compared with 6.7% under accurate estimation. The generated cases illustrate conflicts associated with delayed lane changes and altered passing order. These results characterize controlled, simulator-based stress-test outcomes in the IDM–MOBIL highway-merging setting; they do not constitute empirical validation of human-driver responses to intention misjudgment.

Safe autonomous driving in mixed traffic requires a unified understanding of multimodal interactions and dynamic planning under uncertainty. Existing learning-based methods often fail to capture rare but safety-critical behaviors, while rule-based systems lack adaptability in complex interactions. To address these limitations, we propose CogDrive, a cognition-driven multimodal prediction–planning fusion framework that integrates explicit modal reasoning with safety aware decision optimization. The prediction module introduces cognitive representations of interaction modes based on topological motion semantics and nearest-neighbor relational encoding. By incorporating a differentiable modal loss and multimodal Gaussian decoding, CogDrive effectively learns sparse and unbalanced interaction behaviors, improving long-tail trajectory prediction accuracy. The planning module builds upon an emergency-response concept and develops a safety-stabilized trajectory tree optimization. Short-term consistent root trajectories ensure safety within replanning cycles, while long-term branches provide smooth and collision-free avoidance under low-probability or rapidly switching modes. Experiments on Argoverse2 and INTERACTION datasets show that CogDrive achieves state-of-the-art performance, reducing minADE and miss rate while maintaining smoothness. Closed-loop simulations further confirm stable and adaptive behavior across strong-interaction scenarios such as merging and intersections. By coupling cognitive multimodal prediction with safety-oriented planning, CogDrive establishes an interpretable and reliable paradigm for safe autonomy in complex traffic.

Safe autonomous driving in mixed traffic requires a unified understanding of multimodal interactions and dynamic planning under uncertainty. Existing learning based approaches struggle to capture rare but safety critical behaviors, while rule based systems often lack adaptability in complex interactions. To address these limitations, CogDrive introduces a cognition driven multimodal prediction and planning framework that integrates explicit modal reasoning with safety aware trajectory optimization. The prediction module adopts cognitive representations of interaction modes based on topological motion semantics and nearest neighbor relational encoding. With a differentiable modal loss and multimodal Gaussian decoding, CogDrive learns sparse and unbalanced interaction behaviors and improves long horizon trajectory prediction. The planning module incorporates an emergency response concept and optimizes safety stabilized trajectories, where short term consistent branches ensure safety during replanning cycles and long term branches support smooth and collision free motion under low probability switching modes. Experiments on Argoverse2 and INTERACTION datasets show that CogDrive achieves strong performance in trajectory accuracy and miss rate, while closed loop simulations confirm adaptive behavior in merge and intersection scenarios. By combining cognitive multimodal prediction with safety oriented planning, CogDrive offers an interpretable and reliable paradigm for safe autonomy in complex traffic.

Deep neural networks (DNN) have achieved remarkable success in motion forecasting. However, most DNN-based methods suffer from catastrophic forgetting and fail to maintain their performance in previously learned scenarios after adapting to new data. Recent continual learning (CL) studies aim to mitigate this phenomenon by enhancing memory stability of DNN, i.e., the ability to retain learned knowledge. Yet, excessive emphasis on the memory stability often impairs learning plasticity, i.e., the capacity of DNN to acquire new information effectively. To address such stability-plasticity dilemma, this study proposes a novel CL method, synergetic memory rehearsal (SyReM), for DNN-based motion forecasting. SyReM maintains a compact memory buffer to represent learned knowledge. To ensure memory stability, it employs an inequality constraint that limits increments in the average loss over the memory buffer. Synergistically, a selective memory rehearsal mechanism is designed to enhance learning plasticity by selecting samples from the memory buffer that are most similar to recently observed data. This selection is based on an online-measured cosine similarity of loss gradients, ensuring targeted memory rehearsal. Since replayed samples originate from learned scenarios, this memory rehearsal mechanism avoids compromising memory stability. We validate SyReM under an online CL paradigm where training samples from diverse scenarios arrive as a one-pass stream. Experiments on 11 naturalistic driving datasets from INTERACTION demonstrate that, compared to non-CL and CL baselines, SyReM significantly mitigates catastrophic forgetting in past scenarios while improving forecasting accuracy in new ones. The implementation is publicly available at https://github.com/BIT-Jack/SyReM.

Driving interaction, a critical yet underrepresented element in trajectory datasets, is central to the development and evaluation of autonomous driving systems. This work presents InterHub, a curated dataset of dense multi-agent interaction events, derived from large-scale naturalistic driving recordings. We employ formal methods to describe and extract multi-agent interaction events, exposing the limitations of existing autonomous driving solutions. The dataset is accompanied by an open-source toolkit that enables users to expand InterHub by mining additional interaction events from both public and private driving data. By offering a unified taxonomy, rich annotations, and extensible tools, InterHub supports diverse research needs-from interaction behavior modeling to algorithm benchmarking-and promotes reproducibility, scalability, and cross-dataset comparison in autonomous driving studies.

Decision-making stands as a pivotal component in the realm of autonomous vehicles (AVs), playing a crucial role in navigating the intricacies of autonomous driving. Amidst the evolving landscape of data-driven methodologies, enhancing decision-making performance in complex scenarios has emerged as a prominent research focus. Despite considerable advancements, current learning-based decision-making approaches exhibit potential for refinement, particularly in aspects of policy articulation and safety assurance. To address these challenges, we introduce DDM-Lag, a Diffusion Decision Model, augmented with Lagrangian-based safety enhancements. This work conceptualizes the sequential decision-making challenge inherent in autonomous driving as a problem of generative modeling, adopting diffusion models as the medium for assimilating patterns of decision-making. We introduce a hybrid policy update strategy for diffusion models, amalgamating the principles of behavior cloning and Q-learning, alongside the formulation of an Actor-Critic architecture for the facilitation of updates. To augment the model's exploration process with a layer of safety, we incorporate additional safety constraints, employing a sophisticated policy optimization technique predicated on Lagrangian relaxation to refine the policy learning endeavor comprehensively. Empirical evaluation of our proposed decision-making methodology was conducted across a spectrum of driving tasks, distinguished by their varying degrees of complexity and environmental contexts. The comparative analysis with established baseline methodologies elucidates our model's superior performance, particularly in dimensions of safety and holistic efficacy.

Interacting with other human road users is one of the most challenging tasks for autonomous vehicles. For congruent driving behaviors, it is essential to recognize and comprehend sociality, encompassing both implicit social norms and individualized social preferences of human drivers. To understand and quantify the complex sociality in driving interactions, we propose a Virtual-Game-based Interaction Model (VGIM) that is parameterized by a social preference measurement, Interaction Preference Value (IPV). The IPV is designed to capture the driver's relative inclination towards individual rewards over group rewards. A method for identifying IPV from observed driving trajectory is also developed, with which we assessed human drivers' IPV using driving data recorded in a typical interactive driving scenario, the unprotected left turn. Our findings reveal that (1) human drivers exhibit particular social preference patterns while undertaking specific tasks, such as turning left or proceeding straight; (2) competitive actions could be strategically conducted by human drivers in order to coordinate with others. Finally, we discuss the potential of learning sociality-aware navigation from human demonstrations by incorporating a rule-based humanlike IPV expressing strategy into VGIM and optimization-based motion planners. Simulation experiments demonstrate that (1) IPV identification improves the motion prediction performance in interactive driving scenarios and (2) the dynamic IPV expressing strategy extracted from human driving data makes it possible to reproduce humanlike coordination patterns in the driving interaction.