Selected work

Representative Papers

Importance Sampling and PCA for Finding Failures in Commercial Autonomous Vehicles

Jul 20, 2026

This work addresses the challenge of efficiently uncovering rare yet critical failure scenarios in commercial autonomous driving systems, which are often missed by conventional Monte Carlo simulation. The study proposes a novel approach that integrates adaptive stress testing (AST) with diffusion-based failure sampling (DiFS) to actively search for rare noise trajectories leading to collisions. Furthermore, principal component analysis (PCA) is employed to classify and diagnose distinct failure modes, establishing a closed-loop pipeline from failure discovery to perception defect localization. Evaluated on merging and cut-in scenarios, the method successfully identifies collision cases overlooked by traditional simulation. The extracted canonical noise trajectories consistently reproduce failures across identical or similar scenarios, demonstrating the approach’s effectiveness and transferability.

0 citationsRead paper

Real-Time Rulebook-Aware Nonlinear MPC for Autonomous Driving with Priority-Biased Tiered Slacks

Jul 12, 2026

This work addresses the challenge of autonomous driving motion planning under real-time constraints, requiring a balance among safety, rule compliance, comfort, and efficiency while enabling auditable decision-making. The authors propose W-SQP, a nonlinear model predictive controller based on weighted hierarchical slack variables, which encodes nine categories of driving rules into a four-layer nonlinear program with shared slack variables. A strongly separated hierarchical penalty mechanism prioritizes satisfaction of higher-priority rules while preserving hard actuator constraints. Leveraging CasADi and IPOPT, the system solves the optimization problem online at 10 Hz, guaranteeing feasible solutions at every time step and logging rule residuals for auditability. In closed-loop evaluations across 150 scenarios from the Waymo Open Motion Dataset, the method exhibits no systematic failures in safety or compliance metrics, with only localized performance degradation observed in highly ambiguous, complex scenes.

0 citationsRead paper

Scaling Self-Play for End-to-End Driving

Jun 17, 2026

This work addresses key limitations in end-to-end autonomous driving—namely, reliance on limited human demonstrations, absence of closed-loop feedback, error accumulation, and poor handling of long-tail interactive scenarios—by introducing the first scalable pixel-level self-play training framework. Leveraging the high-throughput Gigapixel simulator for efficient closed-loop learning, the approach integrates a self-play DAgger algorithm, privileged reinforcement learning with teacher policy distillation, and lightweight perception adaptation to enable robust sim-to-real transfer. Experiments demonstrate that the method achieves state-of-the-art performance on the HUGSIM and NAVSIM-v2 benchmarks without any human trajectory supervision, with policy performance scaling linearly with the scale of self-play training.

0 citationsRead paper

Re-imagining ISO 26262 in the Age of Autonomous Vehicles: Enhancing Controllability through Transferability and Predictability

Jun 05, 2026

This study addresses the inadequacy of ISO 26262’s controllability concept—originally defined with respect to human drivers—for SAE Level 4–5 automated driving systems. The authors deconstruct controllability into two quantifiable sub-dimensions: transferability, denoting the system’s ability to hand over control to a fallback mechanism, and predictability, reflecting the extent to which other road users can anticipate the vehicle’s behavior. For the first time, these dimensions are formalized within a rigorous mathematical framework. By analyzing the gap between design intent and practically achievable performance, and integrating human–machine interaction theory, formal modeling, and risk assessment methodologies, the work establishes a falsifiable and traceable set of quantitative metrics. This approach effectively bridges functional safety (ISO 26262) and safety of the intended functionality (SOTIF), enabling the extension of existing standards to high-level automated driving scenarios.

0 citationsRead paper

RECTOR: Priority-Aware Rule-Based Reranking for Compliance-Aware Autonomous Driving Trajectory Selection

May 24, 2026

This work addresses the limitations of existing autonomous driving systems that rely solely on model confidence for trajectory selection, often neglecting multi-level constraints such as safety, traffic regulations, and comfort, thereby leading to frequent violations. To remedy this, the authors propose a hierarchical rulebook-based trajectory reranking method grounded in a priority ordering: safety ≻ legality ≻ road rules ≻ comfort. Their approach integrates differentiable rule proxies and a scene-conditioned applicability mechanism, coupled with a deterministic ε-lexicographic decision policy, enabling provable rule compliance without retraining the underlying prediction model. Evaluated on the Waymo Open Motion Dataset, the method reduces safety and legal violation rates from 28.58% to 20.42% and overall violations from 40.32% to 32.41% compared to confidence-based selection, while maintaining 96% robust rejection capability under adversarial confidence perturbations.

0 citationsRead paper

Recent publications

Latest Papers

Importance Sampling and PCA for Finding Failures in Commercial Autonomous Vehicles

Jul 20, 2026

This work addresses the challenge of efficiently uncovering rare yet critical failure scenarios in commercial autonomous driving systems, which are often missed by conventional Monte Carlo simulation. The study proposes a novel approach that integrates adaptive stress testing (AST) with diffusion-based failure sampling (DiFS) to actively search for rare noise trajectories leading to collisions. Furthermore, principal component analysis (PCA) is employed to classify and diagnose distinct failure modes, establishing a closed-loop pipeline from failure discovery to perception defect localization. Evaluated on merging and cut-in scenarios, the method successfully identifies collision cases overlooked by traditional simulation. The extracted canonical noise trajectories consistently reproduce failures across identical or similar scenarios, demonstrating the approach’s effectiveness and transferability.

0 citationsRead paper

Real-Time Rulebook-Aware Nonlinear MPC for Autonomous Driving with Priority-Biased Tiered Slacks

Jul 12, 2026

This work addresses the challenge of autonomous driving motion planning under real-time constraints, requiring a balance among safety, rule compliance, comfort, and efficiency while enabling auditable decision-making. The authors propose W-SQP, a nonlinear model predictive controller based on weighted hierarchical slack variables, which encodes nine categories of driving rules into a four-layer nonlinear program with shared slack variables. A strongly separated hierarchical penalty mechanism prioritizes satisfaction of higher-priority rules while preserving hard actuator constraints. Leveraging CasADi and IPOPT, the system solves the optimization problem online at 10 Hz, guaranteeing feasible solutions at every time step and logging rule residuals for auditability. In closed-loop evaluations across 150 scenarios from the Waymo Open Motion Dataset, the method exhibits no systematic failures in safety or compliance metrics, with only localized performance degradation observed in highly ambiguous, complex scenes.

0 citationsRead paper

Scaling Self-Play for End-to-End Driving

Jun 17, 2026

This work addresses key limitations in end-to-end autonomous driving—namely, reliance on limited human demonstrations, absence of closed-loop feedback, error accumulation, and poor handling of long-tail interactive scenarios—by introducing the first scalable pixel-level self-play training framework. Leveraging the high-throughput Gigapixel simulator for efficient closed-loop learning, the approach integrates a self-play DAgger algorithm, privileged reinforcement learning with teacher policy distillation, and lightweight perception adaptation to enable robust sim-to-real transfer. Experiments demonstrate that the method achieves state-of-the-art performance on the HUGSIM and NAVSIM-v2 benchmarks without any human trajectory supervision, with policy performance scaling linearly with the scale of self-play training.

0 citationsRead paper

Re-imagining ISO 26262 in the Age of Autonomous Vehicles: Enhancing Controllability through Transferability and Predictability

Jun 05, 2026

This study addresses the inadequacy of ISO 26262’s controllability concept—originally defined with respect to human drivers—for SAE Level 4–5 automated driving systems. The authors deconstruct controllability into two quantifiable sub-dimensions: transferability, denoting the system’s ability to hand over control to a fallback mechanism, and predictability, reflecting the extent to which other road users can anticipate the vehicle’s behavior. For the first time, these dimensions are formalized within a rigorous mathematical framework. By analyzing the gap between design intent and practically achievable performance, and integrating human–machine interaction theory, formal modeling, and risk assessment methodologies, the work establishes a falsifiable and traceable set of quantitative metrics. This approach effectively bridges functional safety (ISO 26262) and safety of the intended functionality (SOTIF), enabling the extension of existing standards to high-level automated driving scenarios.

0 citationsRead paper

RECTOR: Priority-Aware Rule-Based Reranking for Compliance-Aware Autonomous Driving Trajectory Selection

May 24, 2026

This work addresses the limitations of existing autonomous driving systems that rely solely on model confidence for trajectory selection, often neglecting multi-level constraints such as safety, traffic regulations, and comfort, thereby leading to frequent violations. To remedy this, the authors propose a hierarchical rulebook-based trajectory reranking method grounded in a priority ordering: safety ≻ legality ≻ road rules ≻ comfort. Their approach integrates differentiable rule proxies and a scene-conditioned applicability mechanism, coupled with a deterministic ε-lexicographic decision policy, enabling provable rule compliance without retraining the underlying prediction model. Evaluated on the Waymo Open Motion Dataset, the method reduces safety and legal violation rates from 28.58% to 20.42% and overall violations from 40.32% to 32.41% compared to confidence-based selection, while maintaining 96% robust rejection capability under adversarial confidence perturbations.

0 citationsRead paper