UAV Swarm Planning accelerator on FPGA with low latency and fixed-point L-BFGS Quasi-Newton solver
Suquan Zhang, Jincheng Yu, Yuanfan Xu, Yu Wang · 2023
Unmanned aerial vehicle (UAV) swarms have been widely used in search and rescue, surveillance and inspection tasks. Complex scenes, unpredictable environmental obstacles, and mutual interference between multiple aircrafts put forward higher requirements for the robustness and fast calculation of UAV algorithms. Optimization-based planning methods are commonly used for UAV swarms, with the Quasi-Newton method being the most frequently employed solver. However, the QuasiNewton method will generate a huge amount of calculation due to the complex constraints of the environment, which makes real-time planning of UAVs difficult. In order to reduce the time cost of optimization caused by complex constraints, we analyze the dynamic range of data during the calculation of the L-BFGS Quasi-Newton method, and studied the influence of different quantization methods and bit widths on the optimizer operation. We design a hardware accelerator that supports 16 -bit fixed-point numbers and dynamic quantization. The experimental results on Xilinx MPSoc ZU9 show that our implementation achieves a speedup of 8.5 times compared to Intel i9-7920X in Newton’s method calculation with almost no loss of accuracy, and the total power consumption is $1.4 \mathrm{~W}$.