Path planning method for leader-follower multi-vehicle formation with integrating GoT-SAC

doi:10.3969/j.issn.1674-8484.2026.01.013

Abstract

Abstract:

A leader-follower formation path planning method was proposed through integrating the Goal-oriented Transformer (GoT) and the Soft Actor-Critic (SAC) on the Mecanum Wheeled intelligent platform, named GoT-SAC, to enhance the stability and efficiency of formation operation in unknown environments. Experimental validation was conducted in both the Gazebo environment and on a miniature physical platform. The results show that the GoT-SAC model convergences within 95~100 training episodes. The average relative pose error reduces from 18 cm to 6 cm with a path-length relative-difference being below 5% compared with the manual remote-control strategy. Therefore, the proposed method achieves stable formation and efficient obstacle avoidance without relying on prior map information.

Key words: multi-vehicles formation control, leader-follower formation, path planning, deep reinforcement learning

CLC Number:

U467.1+4

WANG Yue, DUAN Hongwei, ZHONG Wei, YANG Lu, HE Lei, CHAI Fulai, SHI Xiaoyang. Path planning method for leader-follower multi-vehicle formation with integrating GoT-SAC[J]. Journal of Automotive Safety and Energy, 2026, 17(1): 122-129.

Figures/Tables 12

References 19

[1]	Balakrishnan S, Azman A D, Nisar J, et al. IoT-enabled smart warehousing with AMR robots and blockchain:A comprehensive approach to efficiency and safety[C]// Proc 3rd Int’l Conf Math Modeling Computational Sci (ICMMCS), Singapore, 2023: 261-270.
[2]	程乐平, 李欢. 智能仓储物流中机器人技术的应用与发展[J]. 信息系统工程, 2023(7): 43-46.
	CHENG Leping, LI Huan. Application and development of robotic technologies in smart warehousing logistics[J]. *J Info Syst Engi*, 2023(7): 43-46. (in Chinese)
[3]	孔国杰, 冯时, 于会龙, 等. 无人集群系统协同运动规划技术综述[J]. 兵工学报, 2023, 44(1): 11-26. doi: 10.12382/bgxb.2022.0930
	KONG Guojie, FENG Shi, YU Huilong, et al. A survey on cooperative motion planning for unmanned swarm systems[J]. *Acta Armament*, 2023, 44(1): 11-26. (in Chinese)
[4]	Cruz J. Leader-follower strategies for multilevel systems[J]. *IEEE Trans Autom Contr*, 1978, 23(2): 244-255. doi: 10.1109/TAC.1978.1101716 URL
[5]	Tan K H, Lewis M A. Virtual structures for high-precision cooperative mobile robotic control[C]// Proc 1996 IEEE/RSJ Int'l Conf Intell Robo Syst (IROS 1996). Osaka, Japan: IEEE, 1996: 132-139.
[6]	Balch T, Arkin R C. Behavior-based formation control for multirobot teams[J]. *IEEE Trans Robot Autom*, 1998, 14(6): 926-939. doi: 10.1109/70.736776 URL
[7]	Olfati-Saber R, Murray RM. Distributed cooperative control of multiple vehicle formations using structural potential functions[C]// Proc IFAC World Congr, Barcelona, Spain: IFAC, 2002: 495-500.
[8]	Costa M M, Silva M F. A survey on path planning algorithms for mobile robots[C]// Proc 2019 IEEE Int’l Conf Autonom Robot Syst Competi (ICARSC). Piscataway: IEEE, 2019: 1-7.
[9]	Hart P E, Nilsson N J, Raphael B. A formal basis for the heuristic determination of minimum cost paths[J]. *IEEE Trans Syst Sci Cybern*, 1968, 4(2): 100-107. doi: 10.1109/TSSC.1968.300136 URL
[10]	Dijkstra E W. A note two problems in connection with graphs[J]. *Numer Math*, 1959, 1(1): 269-271. doi: 10.1007/BF01386390 URL
[11]	劳彩莲, 李鹏, 冯宇. 基于改进A*与DWA算法融合的温室机器人路径规划[J]. 农业机械学报, 2021, 52(1): 14-22.
	LAO Cailian, LI Peng, FENG Yu. Greenhouse robot path planning based on improved A* and DWA fusion[J]. *Trans Chin Soc Agric Engi*, 2021, 52(1): 14-22. (in Chinese)
[12]	郭烈, 齐国栋, 赵一兵, 等. 融合A*与TEB算法的机器人多任务导航调度研究[J]. 华中科技大学学报(自然科学版), 2023, 51(2): 82-88.
	GUO Lie, QI Guodong, ZHAO Yibing, et al. Multi-task navigation scheduling for robots via A and TEB fusion[J]. *J Huazhong Univ of Sci Techn (Nat Sci Edi)*, 2023, 51(2): 82-88. (in Chinese).
[13]	SUN Huihui, ZHANG Weijie, YU Runxiang, et al. Motion planning for mobile robots-Focusing on deep reinforcement learning: A systematic review[J]. *IEEE Access*, 2021, 9: 69061-69081. doi: 10.1109/ACCESS.2021.3076530 URL
[14]	LV Lihua, ZHANG Shujuan, DING Derong, et al. Path planning via an improved DQN-based learning policy[J]. *IEEE Access*, 2019, 7: 67319-67330. doi: 10.1109/ACCESS.2019.2918703
[15]	WANG Shijie, ZHENG Xiang, CAO Yuxiang, et al. A multi-target trajectory planning of a 6-DoF free-floating space robot via reinforcement learning[C]// Int’l Conf Intell Robots Syst (IROS). IEEE, 2021: 3724-3730.
[16]	FENG Zengxi, WANG Chang, AN Jianhu, et al. Emergency fire escape path planning model based on improved DDPG algorithm[J]. *J Build Engi*, 2024, 95: 110090-11011.
[17]	ZHAO Feiyu, LI Dayan, WANG Zhengxu, et al. Autonomous localized path planning algorithm for UAVs based on TD3 strategy[J]. *Sci Rep*, 2024, 14(1): 763-785. doi: 10.1038/s41598-024-51349-4
[18]	GFRERRER A. Geometry and kinematics of the Mecanum wheel[J]. *Comput Aided Geom Des*, 2008, 25(9): 784-791. doi: 10.1016/j.cagd.2008.07.008 URL
[19]	Haastrup A I, Ofuzim O W, Oladejo J A. Kinematic analysis of omnidirectional Mecanum wheeled robot[J]. *Int’l J Engi Appl Phys*, 2023, 3(1): 634-644.

输入：由K帧深度图与目标特征(相对距离、航向误差)经GoT编码得到的场景表示
输出：底盘速度指令
1	用预训练参数φ^*初始化GoT网络
2	初始化SAC算法中的Critic和Actor网络： φ, θ
3	初始化熵参数： α
4	初始化批量大小N并将回放缓冲区N←φ
5	分配目标参数：θ_targ←θ
6	对于episode = 1到E执行：
7	初始化环境状态：S_t~Env
8	初始化目标状态：S_{{goal, t}}~Env
9	对于step = 1到S执行：
10	映射目标标记：G_t = MLP(S_{{goal, t}})
11	场景表示：H_t←GoT(S_t, G_t, φ^*)
12	采样动作：A_t←π_φ(A_t\|A_t)
13	与环境交互：R_t, S_t+1, S_{{goal, t+1}}~Env
14	存储转换：D←D∪(S_t, S_{{goal, t}}, A_t, R_t, S_t+1, S_{{goal, t+1}})
15	更新Critic、Actor与目标网络
16	结束for (episode)

输入：由K帧深度图与目标特征(相对距离、航向误差)经GoT编码得到的场景表示
输出：底盘速度指令
1	用预训练参数φ^*初始化GoT网络
2	初始化SAC算法中的Critic和Actor网络： φ, θ
3	初始化熵参数： α
4	初始化批量大小N并将回放缓冲区N←φ
5	分配目标参数：θ_targ←θ
6	对于episode = 1到E执行：
7	初始化环境状态：S_t~Env
8	初始化目标状态：S_{{goal, t}}~Env
9	对于step = 1到S执行：
10	映射目标标记：G_t = MLP(S_{{goal, t}})
11	场景表示：H_t←GoT(S_t, G_t, φ^*)
12	采样动作：A_t←π_φ(A_t\|A_t)
13	与环境交互：R_t, S_t+1, S_{{goal, t+1}}~Env
14	存储转换：D←D∪(S_t, S_{{goal, t}}, A_t, R_t, S_t+1, S_{{goal, t+1}})
15	更新Critic、Actor与目标网络
16	结束for (episode)

MAX_EPISODES	1 500
LR_A	5×10^-4
LR_C	5×10^-4
TAU	0.001
GAMMA	0.999
AUTO_TUNE	True
ALPHA	1.0
ActorType	GaussianTransformer
CriticType	CNN
TransformerBlocks	3
AttentionHeads	2

MAX_EPISODES	1 500
LR_A	5×10^-4
LR_C	5×10^-4
TAU	0.001
GAMMA	0.999
AUTO_TUNE	True
ALPHA	1.0
ActorType	GaussianTransformer
CriticType	CNN
TransformerBlocks	3
AttentionHeads	2