Research on Indoor Cooperative Positioning Algorithm for Multi-Robot System
Sheng Chen, Hongbo Zhao, Zhijun He, Beihang University, China
Location: Atrium Ballroom
Alternate Number 2
With the development of emerging technologies, robots have entered a new stage of research, and the demand for robots has also increased. Intelligent robots can sense the external environment without a priori information through sensors, realize positioning and navigation, and complete various tasks autonomously. As tasks become more complex, multi-robot systems show the advantages that a single robot can't match. In a complex and dynamic environment, the multi-robot system improves the positioning accuracy of multi-robot systems by sharing and merging various sensor data. When a single robot fails, the multi-robot system can still operate with higher fault tolerance. Considering these advantages, multi-robot co-location has become one of the research hotspots in the field of robotics. This paper has carried out specific research on this.
At present, the traditional filtering algorithms of single robot systems are relatively mature. How to extend these algorithms in multi-robot systems and take advantage of multi-robot systems is a new research direction. In this paper, an algorithm combining EKF (Extended Kalman Filter) and forward neural network is proposed, which is applied in multi-robot system by independent exploration and centralized positioning. Because EKF has large error under nonlinear and noise non-Gaussian conditions, this paper also study a hybrid positioning technique based on EKF and PF(Particle Filter) to improve the accuracy of multi-robot positioning. Finally, the two algorithms are simulated and the simulation results are discussed to meet the theoretical expected results.
In order to apply the two algorithms to the multi-robot system, the hybrid multi-robot system structure is determined and a multi-robot system model is established. Then, the principle and structure of SLAM(Simultaneous Localization and Mapping) are discussed from four aspects: environmental feature extraction, map expression, uncertainty information processing, and SLAM solution.
Our work started with single robot SLAM problem by using EKF. Then aiming at the nonlinearity error of the system, a forward neural network with strong nonlinear mapping ability is introduced to solve the error between model and actual. At the same time, KF is used for weight learning of neural networks to accelerate convergence. Through the combination of the two, the robot SLAM problem is solved better. It is then extended to multi-robot systems. The multi-robot system composed of N robots moves in an unknown two-dimensional environment, adopting independent exploration and centralized positioning. Before the robots meet, the algorithm is used to perform real-time positioning in the respective coordinate systems. After the encounter, the observation information is used to merge with the landmark information.
Another method discussed in this paper is PF+EKF. The PF is discussed and divided into two stages: prediction and update. PF is used to converge the initial condition to the initial error tolerance range of EKF, and then iteratively filtered by EKF to achieve fast and accurate co-localization. After the sensor information of the robot is preprocessed, it obtains its own pose information and relative observation information. Then use EKF and PF technology for secondary integration to achieve coordinated positioning of multiple robots.
In order to verify the effect and accuracy of the two algorithms, a simulation experiment of multi-robot system was established. The preliminary simulation results show that the EKF combined with the forward neural network effectively reduces the error between the mathematical model and the actual model of the system, but it is easy to diverge for systems with strong nonlinearity. Combined with the hybrid positioning technology of EKF and PF, multi-robot can be quickly and accurately co-located under the condition that the initial conditions are unknown or there is a large error. At the end of the thesis, the research work of the full text is summarized, the limitations of the algorithm and simulation of this subject are analyzed, and the next multi-robot research direction is proposed.