Towards Zero-Shot Transfer Across Embodiments For Driving VLAs
研究通过多数据集训练和BEV-Forcing方法解决VLAs模型在自动驾驶中跨数据集和相机设备的零样本迁移问题。
0 citationsRead paper
Selected work
Representative Papers
Future-Interactions-Aware Trajectory Prediction via Braid Theory
Modeling future social interactions efficiently remains a significant challenge in multi-agent trajectory prediction. This work proposes a novel approach that deeply integrates braid theory into the prediction backbone, leveraging a lightweight parallel auxiliary task to accurately model trajectory crossing patterns and thereby enhance the model’s understanding of future interaction intents. Notably, this method achieves substantial improvements in joint trajectory prediction without increasing inference complexity. Experimental results on three standard benchmarks demonstrate clear gains in joint prediction metrics, effectively capturing coordinated behaviors among multiple agents.
0 citationsRead paper
A New Dataset and Framework for Robust Road Surface Classification via Camera-IMU Fusion
Existing road surface classification methods are limited by reliance on a single sensing modality and insufficient environmental diversity in datasets, hindering generalization in complex real-world scenarios. To address this, this work proposes a multimodal approach that fuses camera and IMU data through a lightweight bidirectional cross-attention module and an adaptive gating mechanism, which dynamically adjusts modality contributions to mitigate domain shift. We introduce ROAD, the first multimodal dataset comprising real-world, vision-dominant, and synthetic subsets, with synchronized RGB-IMU capture. Experimental results demonstrate that our method improves performance by 1.4 percentage points on the PVS benchmark and by 11.6 percentage points on the ROAD multimodal subset, while maintaining high F1 scores under challenging conditions such as nighttime and heavy rain.
0 citationsRead paper
Improving Consistency in Vehicle Trajectory Prediction Through Preference Optimization
Existing deep learning-based trajectory prediction models struggle to capture strong inter-agent dependencies in complex interactive scenarios, leading to inconsistent predictions that compromise autonomous driving safety. To address this, we propose the first preference optimization framework tailored for vehicle trajectory prediction: it incorporates human behavioral priors—automatically derived from relative preferences over future trajectory sequences (e.g., collision avoidance, social plausibility, and motion smoothness)—into multi-agent joint prediction, thereby enhancing consistency without additional inference overhead. Our method requires no architectural modifications; instead, it achieves improved cooperative rationality via preference-driven fine-tuning alone. Evaluated on three benchmark datasets—nuScenes, Argoverse 2, and INTERACTION—it significantly improves scene consistency metrics (average +12.7%) while maintaining state-of-the-art trajectory accuracy (ADE/FDE remain virtually unchanged), demonstrating both effectiveness and practicality.
0 citationsRead paper
Evaluating link prediction: New perspectives and recommendations
Existing link prediction (LP) evaluation lacks systematic control over critical factors—including network type, geodesic distance distribution, class imbalance, and metric sensitivity—limiting the generalizability of empirical conclusions. Method: We propose the first hypothesis-driven, multidimensional controllable evaluation framework, employing controlled-variable experiments, multi-network benchmarking, and rigorous statistical testing to systematically identify and quantify six previously overlooked sources of evaluation bias. Contribution/Results: We reveal the substantial impact of geodesic distance distribution and class imbalance on the performance of mainstream LP methods; demonstrate the inadequacy of conventional metrics (e.g., AUC) in early-retrieval scenarios; and introduce a hierarchical evaluation paradigm alongside application-oriented best-practice guidelines. This work establishes a methodological foundation for fair, reproducible comparison and reliable deployment of LP methods.
0 citationsRead paper
Recent publications
Latest Papers
Towards Zero-Shot Transfer Across Embodiments For Driving VLAs
研究通过多数据集训练和BEV-Forcing方法解决VLAs模型在自动驾驶中跨数据集和相机设备的零样本迁移问题。
0 citationsRead paper
Future-Interactions-Aware Trajectory Prediction via Braid Theory
Modeling future social interactions efficiently remains a significant challenge in multi-agent trajectory prediction. This work proposes a novel approach that deeply integrates braid theory into the prediction backbone, leveraging a lightweight parallel auxiliary task to accurately model trajectory crossing patterns and thereby enhance the model’s understanding of future interaction intents. Notably, this method achieves substantial improvements in joint trajectory prediction without increasing inference complexity. Experimental results on three standard benchmarks demonstrate clear gains in joint prediction metrics, effectively capturing coordinated behaviors among multiple agents.
0 citationsRead paper
A New Dataset and Framework for Robust Road Surface Classification via Camera-IMU Fusion
Existing road surface classification methods are limited by reliance on a single sensing modality and insufficient environmental diversity in datasets, hindering generalization in complex real-world scenarios. To address this, this work proposes a multimodal approach that fuses camera and IMU data through a lightweight bidirectional cross-attention module and an adaptive gating mechanism, which dynamically adjusts modality contributions to mitigate domain shift. We introduce ROAD, the first multimodal dataset comprising real-world, vision-dominant, and synthetic subsets, with synchronized RGB-IMU capture. Experimental results demonstrate that our method improves performance by 1.4 percentage points on the PVS benchmark and by 11.6 percentage points on the ROAD multimodal subset, while maintaining high F1 scores under challenging conditions such as nighttime and heavy rain.
0 citationsRead paper
Improving Consistency in Vehicle Trajectory Prediction Through Preference Optimization
Existing deep learning-based trajectory prediction models struggle to capture strong inter-agent dependencies in complex interactive scenarios, leading to inconsistent predictions that compromise autonomous driving safety. To address this, we propose the first preference optimization framework tailored for vehicle trajectory prediction: it incorporates human behavioral priors—automatically derived from relative preferences over future trajectory sequences (e.g., collision avoidance, social plausibility, and motion smoothness)—into multi-agent joint prediction, thereby enhancing consistency without additional inference overhead. Our method requires no architectural modifications; instead, it achieves improved cooperative rationality via preference-driven fine-tuning alone. Evaluated on three benchmark datasets—nuScenes, Argoverse 2, and INTERACTION—it significantly improves scene consistency metrics (average +12.7%) while maintaining state-of-the-art trajectory accuracy (ADE/FDE remain virtually unchanged), demonstrating both effectiveness and practicality.
0 citationsRead paper
Evaluating link prediction: New perspectives and recommendations
Existing link prediction (LP) evaluation lacks systematic control over critical factors—including network type, geodesic distance distribution, class imbalance, and metric sensitivity—limiting the generalizability of empirical conclusions. Method: We propose the first hypothesis-driven, multidimensional controllable evaluation framework, employing controlled-variable experiments, multi-network benchmarking, and rigorous statistical testing to systematically identify and quantify six previously overlooked sources of evaluation bias. Contribution/Results: We reveal the substantial impact of geodesic distance distribution and class imbalance on the performance of mainstream LP methods; demonstrate the inadequacy of conventional metrics (e.g., AUC) in early-retrieval scenarios; and introduce a hierarchical evaluation paradigm alongside application-oriented best-practice guidelines. This work establishes a methodological foundation for fair, reproducible comparison and reliable deployment of LP methods.
0 citationsRead paper