Selected work

Representative Papers

Explainable deep learning improves human mental models of self-driving cars

Nov 27, 2024arXiv.org

The opacity of black-box motion planners in autonomous driving severely undermines human–machine collaborative safety, while existing eXplainable AI (XAI) methods remain largely confined to simulation or simplified scenarios, lacking real-road validation. To address this, we propose the Concept-Wrapping Network (CW-Net), the first approach enabling causally faithful, performance-preserving, and human-interpretable decision explanations within production-grade autonomous driving systems. CW-Net maps neural network outputs onto a semantically grounded driving concept space—e.g., “yielding” or “emergency evasive maneuver”—by jointly integrating causal reasoning and human cognitive modeling. Real-world deployment evaluations demonstrate that CW-Net significantly improves drivers’ prediction accuracy of vehicle behavior (+28.6%) and enhances response adaptability. This work establishes the first practically deployable, explanation-aware motion planning paradigm for trustworthy human–autonomy collaboration.

1 citationsRead paper

ControlMap: Controllable High-Definition Map Generation for Traffic Scenario Simulation

Jun 14, 2026

This work addresses key limitations in current autonomous driving simulation—namely, the high cost of high-definition (HD) map generation, insufficient scenario diversity, and weak controllability over road topology. To overcome these challenges, we propose a spatially conditioned framework based on latent diffusion models integrated with ControlNet, which, for the first time, introduces spatial guidance signals into diffusion-based HD map synthesis. Our approach enables generation conditioned on specified road topologies, supports urban style transfer, and allows fine-grained control over guidance strength. We further introduce two novel metrics to evaluate both adherence to input constraints and map realism, and demonstrate city-scale style modeling. Experiments show that our method generates maps that accurately follow prescribed topologies while preserving distinctive urban details, significantly enhancing both the diversity and realism of simulated driving scenarios.

0 citationsRead paper

VISA: VLM-Guided Instance Semantic Auditing for 3D Occupancy World Models

Jun 11, 2026

This work addresses the adverse impact of semantic errors—particularly for rare object categories—on free space estimation, collision detection, and temporal consistency in 3D semantic occupancy models. To mitigate this, the authors propose VISA, a training-time semantic auditing framework that leverages a vision-language model (VLM) not for embedding alignment but as a reliability-aware auditor. VISA generates structured semantic audits for each object instance, encompassing category hypotheses, confusions, attributes, and supporting evidence, which are propagated along trajectories to corresponding 3D voxels. A reliability-weighted distillation loss then transfers this knowledge to semantic logits, eliminating the need for VLM inference at test time. Evaluated on nuScenes, VISA improves the mIoU of OccWorld and GaussianWorld to 20.05 and 21.91, respectively, with GaussianWorld achieving 19.16 mIoU on objects and 16.79 on rare categories.

0 citationsRead paper

nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving

May 29, 2026

This work addresses the lack of effective supervision for commonsense reasoning, spatial relationships, and decision inference in long-tail scenarios within existing autonomous driving datasets. The authors introduce a large-scale, real-world driving dataset and benchmark specifically targeting long-tail situations, comprising 20,000 multimodal driving clips. For the first time in real-world data, they systematically incorporate three types of human-verified reasoning annotations—spatial, decision-based, and counterfactual. By integrating multi-camera inputs, LiDAR, high-definition maps, and object annotations, the framework enables joint training and evaluation of reasoning and planning through vision-language models (VLMs) and vision-language-action models (VLAs). Experiments demonstrate that VLMs fine-tuned on this dataset significantly improve driving-related question-answering accuracy, while VLAs trained with reasoning supervision enhance planning performance even when reasoning outputs are disabled.

0 citationsRead paper

SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model

Apr 21, 2026

This work addresses the high inference latency and insufficient robustness of existing vision-language-action (VLA) models in autonomous driving, particularly their difficulty in handling long-tailed and complex reasoning scenarios. The authors propose SpanVLA, a novel framework that integrates an autoregressive vision-language model with a flow-matching action expert. SpanVLA efficiently generates future trajectories through a flow-matching strategy initialized by historical trajectories and introduces, for the first time, a joint learning mechanism combining negative samples and recovery behaviors. Leveraging GRPO post-training and a newly curated mReasoning dataset, SpanVLA significantly reduces inference latency while achieving superior planning performance and robustness on the NAVSIM v1 and v2 benchmarks.

0 citationsRead paper

Recent publications

Latest Papers

ControlMap: Controllable High-Definition Map Generation for Traffic Scenario Simulation

Jun 14, 2026

This work addresses key limitations in current autonomous driving simulation—namely, the high cost of high-definition (HD) map generation, insufficient scenario diversity, and weak controllability over road topology. To overcome these challenges, we propose a spatially conditioned framework based on latent diffusion models integrated with ControlNet, which, for the first time, introduces spatial guidance signals into diffusion-based HD map synthesis. Our approach enables generation conditioned on specified road topologies, supports urban style transfer, and allows fine-grained control over guidance strength. We further introduce two novel metrics to evaluate both adherence to input constraints and map realism, and demonstrate city-scale style modeling. Experiments show that our method generates maps that accurately follow prescribed topologies while preserving distinctive urban details, significantly enhancing both the diversity and realism of simulated driving scenarios.

0 citationsRead paper

VISA: VLM-Guided Instance Semantic Auditing for 3D Occupancy World Models

Jun 11, 2026

This work addresses the adverse impact of semantic errors—particularly for rare object categories—on free space estimation, collision detection, and temporal consistency in 3D semantic occupancy models. To mitigate this, the authors propose VISA, a training-time semantic auditing framework that leverages a vision-language model (VLM) not for embedding alignment but as a reliability-aware auditor. VISA generates structured semantic audits for each object instance, encompassing category hypotheses, confusions, attributes, and supporting evidence, which are propagated along trajectories to corresponding 3D voxels. A reliability-weighted distillation loss then transfers this knowledge to semantic logits, eliminating the need for VLM inference at test time. Evaluated on nuScenes, VISA improves the mIoU of OccWorld and GaussianWorld to 20.05 and 21.91, respectively, with GaussianWorld achieving 19.16 mIoU on objects and 16.79 on rare categories.

0 citationsRead paper

nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving

May 29, 2026

This work addresses the lack of effective supervision for commonsense reasoning, spatial relationships, and decision inference in long-tail scenarios within existing autonomous driving datasets. The authors introduce a large-scale, real-world driving dataset and benchmark specifically targeting long-tail situations, comprising 20,000 multimodal driving clips. For the first time in real-world data, they systematically incorporate three types of human-verified reasoning annotations—spatial, decision-based, and counterfactual. By integrating multi-camera inputs, LiDAR, high-definition maps, and object annotations, the framework enables joint training and evaluation of reasoning and planning through vision-language models (VLMs) and vision-language-action models (VLAs). Experiments demonstrate that VLMs fine-tuned on this dataset significantly improve driving-related question-answering accuracy, while VLAs trained with reasoning supervision enhance planning performance even when reasoning outputs are disabled.

0 citationsRead paper

SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model

Apr 21, 2026

This work addresses the high inference latency and insufficient robustness of existing vision-language-action (VLA) models in autonomous driving, particularly their difficulty in handling long-tailed and complex reasoning scenarios. The authors propose SpanVLA, a novel framework that integrates an autoregressive vision-language model with a flow-matching action expert. SpanVLA efficiently generates future trajectories through a flow-matching strategy initialized by historical trajectories and introduces, for the first time, a joint learning mechanism combining negative samples and recovery behaviors. Leveraging GRPO post-training and a newly curated mReasoning dataset, SpanVLA significantly reduces inference latency while achieving superior planning performance and robustness on the NAVSIM v1 and v2 benchmarks.

0 citationsRead paper

TreeIRL: Safe Urban Driving with Tree Search and Inverse Reinforcement Learning

Sep 16, 2025

Addressing the challenge of simultaneously ensuring safety, efficiency, and human-like driving behavior in urban autonomous driving, this paper proposes a novel trajectory planning framework integrating Monte Carlo Tree Search (MCTS) and deep Inverse Reinforcement Learning (IRL). Methodologically, it is the first to apply MCTS to open-road autonomous driving planning, generating multiple safe candidate trajectories; subsequently, deep IRL infers human driving preferences from expert demonstrations to score and select trajectories based on human-likeness. The framework achieves joint optimization across safety, ride comfort, and behavioral similarity to human drivers. Extensive real-world validation—over 500 miles in Las Vegas urban environments—demonstrates robust performance, while large-scale simulation benchmarks show significant improvements over both conventional and state-of-the-art planners. The proposed approach establishes new performance benchmarks, achieving state-of-the-art (SOTA) results in comprehensive evaluation metrics.

0 citationsRead paper