Selected work

Representative Papers

GuideFlow: Constraint-Guided Flow Matching for Planning in End-to-End Autonomous Driving

Nov 23, 2025

In end-to-end autonomous driving planning, imitation-based methods suffer from multimodal trajectory collapse, while generative approaches struggle to incorporate safety and physical constraints. This paper proposes a constraint-guided flow matching framework: it is the first to explicitly integrate safety and physical constraints directly into the flow matching process; jointly trains an energy-based model (EBM) to enhance autonomous optimization; and introduces driving aggressiveness as a controllable conditional signal to enable diverse, regulation-compliant, and style-tunable trajectory generation. The method achieves state-of-the-art performance across multiple benchmarks—including Bench2Drive, nuScenes, NavSim, and ADV-nuScenes—attaining an EPDMS score of 43.0 on the challenging Navhard test set. Key contributions include: (1) joint modeling of constraints and flow matching, (2) EBM-augmented collaborative optimization, and (3) a controllable trajectory generation mechanism.

0 citationsRead paper

Beyond Imitation: Constraint-Aware Trajectory Generation with Flow Matching For End-to-End Autonomous Driving

Oct 30, 2025

End-to-end autonomous driving planning faces two key challenges: (1) imitation learning often suffers from mode collapse, limiting trajectory diversity; and (2) existing generative models lack explicit integration of safety and kinematic constraints during sampling, necessitating post-hoc correction. To address these, we propose Constraint-Aware Trajectory Generation (CATG), the first framework to embed explicit safety constraints—such as collision avoidance and lane keeping—as well as vehicle kinematic models directly into the flow matching process. CATG further introduces tunable driving aggressiveness as a conditional control signal, enabling controllable, regulation-compliant, and diverse trajectory generation. By leveraging multi-condition guidance and constraint-augmented flow matching, CATG achieves second place (EPDMS 51.31) and the Innovation Award in the NavSim v2 Challenge.

0 citationsRead paper

Expanding-and-Shrinking Binary Neural Networks

Mar 31, 2025

Binary neural networks (BNNs) suffer from limited feature representation capacity due to binary weights and activations, resulting in substantially lower accuracy than full-precision models on complex tasks. To address this, we propose a differentiable, lightweight Expand-Squeeze operation—the first method to break the feature map value bottleneck in BNNs without increasing binary hardware overhead. Our approach enhances binarization via gradient approximation and channel-wise adaptive scaling, thereby improving feature diversity. The operation is architecture-agnostic, seamlessly integrating into both CNNs and Transformers, and compatible with standard BNN training pipelines. Extensive experiments demonstrate state-of-the-art performance across diverse tasks—including image classification, object detection, and diffusion models—achieving absolute accuracy gains of 3.2–5.7 percentage points over prior methods, while incurring less than 0.5% additional computational cost.

0 citationsRead paper

Recent publications

Latest Papers

GuideFlow: Constraint-Guided Flow Matching for Planning in End-to-End Autonomous Driving

Nov 23, 2025

In end-to-end autonomous driving planning, imitation-based methods suffer from multimodal trajectory collapse, while generative approaches struggle to incorporate safety and physical constraints. This paper proposes a constraint-guided flow matching framework: it is the first to explicitly integrate safety and physical constraints directly into the flow matching process; jointly trains an energy-based model (EBM) to enhance autonomous optimization; and introduces driving aggressiveness as a controllable conditional signal to enable diverse, regulation-compliant, and style-tunable trajectory generation. The method achieves state-of-the-art performance across multiple benchmarks—including Bench2Drive, nuScenes, NavSim, and ADV-nuScenes—attaining an EPDMS score of 43.0 on the challenging Navhard test set. Key contributions include: (1) joint modeling of constraints and flow matching, (2) EBM-augmented collaborative optimization, and (3) a controllable trajectory generation mechanism.

0 citationsRead paper

Beyond Imitation: Constraint-Aware Trajectory Generation with Flow Matching For End-to-End Autonomous Driving

Oct 30, 2025

End-to-end autonomous driving planning faces two key challenges: (1) imitation learning often suffers from mode collapse, limiting trajectory diversity; and (2) existing generative models lack explicit integration of safety and kinematic constraints during sampling, necessitating post-hoc correction. To address these, we propose Constraint-Aware Trajectory Generation (CATG), the first framework to embed explicit safety constraints—such as collision avoidance and lane keeping—as well as vehicle kinematic models directly into the flow matching process. CATG further introduces tunable driving aggressiveness as a conditional control signal, enabling controllable, regulation-compliant, and diverse trajectory generation. By leveraging multi-condition guidance and constraint-augmented flow matching, CATG achieves second place (EPDMS 51.31) and the Innovation Award in the NavSim v2 Challenge.

0 citationsRead paper

Expanding-and-Shrinking Binary Neural Networks

Mar 31, 2025

Binary neural networks (BNNs) suffer from limited feature representation capacity due to binary weights and activations, resulting in substantially lower accuracy than full-precision models on complex tasks. To address this, we propose a differentiable, lightweight Expand-Squeeze operation—the first method to break the feature map value bottleneck in BNNs without increasing binary hardware overhead. Our approach enhances binarization via gradient approximation and channel-wise adaptive scaling, thereby improving feature diversity. The operation is architecture-agnostic, seamlessly integrating into both CNNs and Transformers, and compatible with standard BNN training pipelines. Extensive experiments demonstrate state-of-the-art performance across diverse tasks—including image classification, object detection, and diffusion models—achieving absolute accuracy gains of 3.2–5.7 percentage points over prior methods, while incurring less than 0.5% additional computational cost.

0 citationsRead paper