Cascaded Multitask Convolutional Network for Robot Formation

Robot formation control greatly relies on the accuracy and real-time performance associated with acquiring the information of the leader robot. The traditional communication and vision-based methods lead to a large delay and lack robustness to noise. In order to satisfy both requirements, we propose a cascaded multitask convolutional network to jointly address target detection and key point detection. In order to achieve high flexibility, we perform experiments using different model hyper parameters and explore the trade-off between accuracy and real-time performance. The experimental results demonstrate the effectiveness of our method for acquiring the information of the leader robot in real time with high accuracy. Furthermore, our method can be easily adapted to other vision-based tasks, laying foundation for the design of vision-based controllers for robots.