OPUS-V2: Bridging the Gap between Sparse Points and Dense Voxels
为解决点云预测与自驾车系统所需密集体素不匹配的问题，提出OPUS-V2框架，通过轻量级点-体素转换模块提高模型准确性和适应性。
0 citationsRead paper
Selected work
Representative Papers
nuTruck: Benchmarking Autonomous Driving Planning for Distributed Electric-drive Trucks
Existing learning-based planning methods for distributed electric-drive heavy trucks (DETs) lack a high-fidelity closed-loop evaluation benchmark. This work proposes nuTruck—the first autonomous planning benchmark specifically designed for DETs—featuring a high-accuracy nonlinear vehicle dynamics model that supports independent control of all wheels for both driving and steering, and enables quantitative assessment of rollover risk. The platform is compatible with both learning-based and rule-based planners, facilitating large-scale closed-loop simulation and data-driven evaluation under realistic driving scenarios. Beyond verifying collision-free performance, nuTruck introduces, for the first time, a systematic quantification of trajectory dynamic safety, thereby establishing a new standard for autonomous planning in DETs.
0 citationsRead paper
Physics-informed Deep Mixture-of-Koopmans Vehicle Dynamics Model with Dual-branch Encoder for Distributed Electric-drive Trucks
This work addresses the limitations of conventional modeling approaches for distributed electric-drive trucks, which struggle with efficiency, accuracy, and compatibility with linear control due to strong nonlinearities and longitudinal–lateral coupling. To overcome these challenges, the authors propose a fully data-driven dynamics modeling framework grounded in Koopman operator theory. By employing a dual-branch encoder architecture and a geometry-consistent physics-informed supervision mechanism, the method maps the nonlinear system into a linear embedding space. A hybrid Koopman framework is further introduced to accommodate multiple driving modes. Validated through high-fidelity TruckSim simulations and real-vehicle experiments, the approach achieves state-of-the-art long-term state estimation accuracy while maintaining high precision, interpretability, and control-friendly properties.
0 citationsRead paper
MetaDAT: Generalizable Trajectory Prediction via Meta Pre-training and Data-Adaptive Test-Time Updating
This work addresses the performance degradation of trajectory prediction models during testing due to distribution shifts by proposing a meta-learning pretraining framework tailored for trajectory forecasting, which integrates test-time training (TTT) with a data-adaptive updating mechanism. The approach employs bi-level optimization during meta-pretraining to enhance the model’s rapid adaptation capability. At test time, it dynamically adjusts the learning rate and update frequency, while leveraging hard example mining and online gradient analysis to focus on critical samples, enabling efficient and accurate online adaptation. Evaluated across diverse datasets—including nuScenes, Lyft, and Waymo—the method significantly outperforms existing approaches and demonstrates robustness and high performance even under suboptimal learning rates or high frame rates.
0 citationsRead paper
Online Segment Any 3D Thing as Instance Tracking
Existing online 3D segmentation methods rely on predefined object queries and neglect temporal dynamics in perception, rendering them vulnerable to viewpoint changes, occlusions, and fragmented features from vision foundation models (VFMs), thereby yielding incoherent instance associations and limited holistic understanding. This work pioneers modeling online 3D segmentation as a spatiotemporal instance tracking task. We propose a sparse object query-based framework for temporal feature propagation, integrating long-range instance association with short-term observation refinement, and introduce spatiotemporal consistency learning to mitigate occlusion and feature fragmentation. Our method operates directly on VFM-driven 3D point clouds without requiring additional annotations. On ScanNet200, it achieves a +2.8 AP gain over ESAM; consistent improvements are observed across ScanNet, SceneNN, and 3RScan. The approach significantly enhances fine-grained, temporally coherent instance perception in dynamic environments.
0 citationsRead paper
Recent publications
Latest Papers
OPUS-V2: Bridging the Gap between Sparse Points and Dense Voxels
为解决点云预测与自驾车系统所需密集体素不匹配的问题，提出OPUS-V2框架，通过轻量级点-体素转换模块提高模型准确性和适应性。
0 citationsRead paper
nuTruck: Benchmarking Autonomous Driving Planning for Distributed Electric-drive Trucks
Existing learning-based planning methods for distributed electric-drive heavy trucks (DETs) lack a high-fidelity closed-loop evaluation benchmark. This work proposes nuTruck—the first autonomous planning benchmark specifically designed for DETs—featuring a high-accuracy nonlinear vehicle dynamics model that supports independent control of all wheels for both driving and steering, and enables quantitative assessment of rollover risk. The platform is compatible with both learning-based and rule-based planners, facilitating large-scale closed-loop simulation and data-driven evaluation under realistic driving scenarios. Beyond verifying collision-free performance, nuTruck introduces, for the first time, a systematic quantification of trajectory dynamic safety, thereby establishing a new standard for autonomous planning in DETs.
0 citationsRead paper
Physics-informed Deep Mixture-of-Koopmans Vehicle Dynamics Model with Dual-branch Encoder for Distributed Electric-drive Trucks
This work addresses the limitations of conventional modeling approaches for distributed electric-drive trucks, which struggle with efficiency, accuracy, and compatibility with linear control due to strong nonlinearities and longitudinal–lateral coupling. To overcome these challenges, the authors propose a fully data-driven dynamics modeling framework grounded in Koopman operator theory. By employing a dual-branch encoder architecture and a geometry-consistent physics-informed supervision mechanism, the method maps the nonlinear system into a linear embedding space. A hybrid Koopman framework is further introduced to accommodate multiple driving modes. Validated through high-fidelity TruckSim simulations and real-vehicle experiments, the approach achieves state-of-the-art long-term state estimation accuracy while maintaining high precision, interpretability, and control-friendly properties.
0 citationsRead paper
MetaDAT: Generalizable Trajectory Prediction via Meta Pre-training and Data-Adaptive Test-Time Updating
This work addresses the performance degradation of trajectory prediction models during testing due to distribution shifts by proposing a meta-learning pretraining framework tailored for trajectory forecasting, which integrates test-time training (TTT) with a data-adaptive updating mechanism. The approach employs bi-level optimization during meta-pretraining to enhance the model’s rapid adaptation capability. At test time, it dynamically adjusts the learning rate and update frequency, while leveraging hard example mining and online gradient analysis to focus on critical samples, enabling efficient and accurate online adaptation. Evaluated across diverse datasets—including nuScenes, Lyft, and Waymo—the method significantly outperforms existing approaches and demonstrates robustness and high performance even under suboptimal learning rates or high frame rates.
0 citationsRead paper
Online Segment Any 3D Thing as Instance Tracking
Existing online 3D segmentation methods rely on predefined object queries and neglect temporal dynamics in perception, rendering them vulnerable to viewpoint changes, occlusions, and fragmented features from vision foundation models (VFMs), thereby yielding incoherent instance associations and limited holistic understanding. This work pioneers modeling online 3D segmentation as a spatiotemporal instance tracking task. We propose a sparse object query-based framework for temporal feature propagation, integrating long-range instance association with short-term observation refinement, and introduce spatiotemporal consistency learning to mitigate occlusion and feature fragmentation. Our method operates directly on VFM-driven 3D point clouds without requiring additional annotations. On ScanNet200, it achieves a +2.8 AP gain over ESAM; consistent improvements are observed across ScanNet, SceneNN, and 3RScan. The approach significantly enhances fine-grained, temporally coherent instance perception in dynamic environments.
0 citationsRead paper