PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference
为解决视觉语言模型推理成本高和性能下降问题，提出PACE框架，通过压缩和提取范式加速视觉编码器和大语言模型的推理。
0 citationsRead paper
Selected work
Representative Papers
Optimal Smooth Coverage Trajectory Planning for Quadrotors in Cluttered Environment
This paper addresses the trajectory planning problem for unmanned aerial vehicles (UAVs) in complex scenarios such as power line inspection, where full coverage of points of interest (POIs) must be achieved under smoothness, timeliness, and obstacle-avoidance constraints. We propose a front-end/back-end co-design framework for optimal smooth trajectory generation. The front-end employs a genetic algorithm to solve the POI visiting sequence, formulated as a constrained traveling salesman problem (TSP). The back-end jointly optimizes trajectory timing, C² continuity, and minimum obstacle clearance via nonlinear least-squares, yielding differentiable, safe, and time-efficient fully covered paths. Our key innovation lies in the tight integration of discrete sequence optimization with continuous trajectory synthesis, augmented by environment-aware explicit obstacle modeling. Numerical simulations demonstrate that the algorithm robustly generates smooth, 100% POI-covered trajectories in dense obstacle environments, reducing average mission time by 18.7% while maintaining feasibility—indicating strong potential for real-time deployment.
0 citationsRead paper
Recent publications
Latest Papers
PACE: A Unified Condense-and-Extract Paradigm for Fast VLM Inference
为解决视觉语言模型推理成本高和性能下降问题，提出PACE框架，通过压缩和提取范式加速视觉编码器和大语言模型的推理。
0 citationsRead paper
Optimal Smooth Coverage Trajectory Planning for Quadrotors in Cluttered Environment
This paper addresses the trajectory planning problem for unmanned aerial vehicles (UAVs) in complex scenarios such as power line inspection, where full coverage of points of interest (POIs) must be achieved under smoothness, timeliness, and obstacle-avoidance constraints. We propose a front-end/back-end co-design framework for optimal smooth trajectory generation. The front-end employs a genetic algorithm to solve the POI visiting sequence, formulated as a constrained traveling salesman problem (TSP). The back-end jointly optimizes trajectory timing, C² continuity, and minimum obstacle clearance via nonlinear least-squares, yielding differentiable, safe, and time-efficient fully covered paths. Our key innovation lies in the tight integration of discrete sequence optimization with continuous trajectory synthesis, augmented by environment-aware explicit obstacle modeling. Numerical simulations demonstrate that the algorithm robustly generates smooth, 100% POI-covered trajectories in dense obstacle environments, reducing average mission time by 18.7% while maintaining feasibility—indicating strong potential for real-time deployment.
0 citationsRead paper