Reliable Control of a Quadruped Walking Robot in Uneven Terrain Environment based on Noninvasive Brain-Computer Interface 1 Wenchuan Jia and Huayan Pu Xin Luo and Xuedong Chen* Dandan Huang and Ou Bai School of Mechatronic Engineering and Automation State Key Laboratory of Digital Manufacturing Equipment and Technology EEG&BCI Laboratory, Department of Biomedical Engineering Shanghai University Huazhong University of Science&Technology Virginia Commonwealth University Shanghai, P. R. China Wuhan, P. R. China Richmond, Virginia, US. {lovvchris&phygood_2001}@shu.edu.cn {chenxd & mexinluo}@mail. hust.edu.cn {obai & huangd2}@vcu.edu This work was supported by the National Natural Science Foundation of China under grant 51121002. *Corresponding author: Xuedong Chen (Tel: +86-27-87557325; fax: +86-27-87557325; e-mail: chenxd@mail.hust.edu.cn) Abstract - In this paper, a novel control architecture and strategies, involving noninvasive brain-computer interface (BCI), are presented, and a virtual prototype of quadruped walking robot, named QWR-I, is developed for simulation of navigation behaviors of the BCI-based robot. The BCI can provide limited patterns of user’s intention based on EEG signals, following user’s motor imagination. The proposed control architecture, dealing with both robot’s autonomous planning and user’s decision acquired from BCI, is elaborated. To achieve efficient navigation in uneven terrain environment where the robot is located, three control modes are proposed, taking trade-off between robot’s autonomy and user’s flexibility. Movement efficiency is priority in relatively even terrain by strolling mode, while safety is considered more in uneven terrain by terrain mode. The coding protocols, which semantically represent the same set of BCI signals with different meanings according to the situated control mode of the robot, and switching strategies between different modes are elucidated. To satisfy the user's intention of desired path to destination to greatest extent possible on the premise of walking stability, cooperative control strategy based on centroid stability margin while the robot walks is presented. Keywords: Cognitive human-robot interaction, quadruped walking, cooperative navigation, brain-control I. INTRODUCTION Brain activity can be captured invasively by the electrodes implanted in the human brain or non-invasively by the electrodes placed on the scalp [1]. For users, the non-invasive methods with low clinical threshold are more acceptable, although the signal to noise ratio is lower compared with the invasive methods due to the electrical signals dampened and blurred when picked up from outside the skull. In practical applications, electroencephalography (EEG)-based BCI has been adopted to interpret user’s intention. Studies undertaken by Iturrate et al. [2] show that subjects could achieve mental control to drive an intelligent wheelchair with high robustness and low variability. Millán et al. [3] presents asynchronous and non-invasive BCI for continuous control of intelligent wheelchair. Some other EEG-based robotic systems have also been developed [4]-[8]. Nevertheless, the mobility is restricted as wheelchairs can only maneuver around terrain suitable for wheeled vehicle. There are various terrain structures in real environment, from flat sidewalks to rocky mountain trails. Based on the environmental condition, people navigate with different manners. Therefore, prior to navigation in variable terrain environment, different walking manners should be identified and investigated. Different walking manners, also called walking modes, are certain navigation methods for a specific type of terrain. The user has to understand each mode and learn how to switch among them. Since most people have difficulty even understanding the modal operations of programming a VHS, adopting many modes with numerous mode transition commands is extremely undesirable. To equip brain-controlled walking robot with all terrain mobility, a critical problem that how to design the execution protocol using limited types of EEG signals which has insufficient decoding accuracy and time delay should be solved [7]-[8]. The user can specify his/her control intention to interact with the robot based on the execution protocol. The walking process done with a group of specific execution protocols makes a large range of potential actions capable of being executed by the user. As mentioned above, a minimal amount of user input is desired, so correlations depending on the environment and the user’s actions will aid in simplifying the required actions. However, the correlations create a difficult tradeoff for each execution of which the implementation is complex. Execution protocol based on few intention commands requires robot to have a large preprogrammed bank of intelligent correlations between user inputs and desired actions. Simple execution protocol based on user’s desires will be discussed in control system design. II. HUMAN-ROBOT COOPERATION SYSTEM A. EEG Signal Different types of EEG signals have different extraction accuracy. In this study, four types of EEG signals which have relatively high spatiotemporal resolutions [9] are adopted for BCI, i.e. Idle, Lno, Rno and LRyes. All the signals are expected to be obtained through performing corresponding imagination of hand movement tasks. The details can be found in [9]. We apply signal processing and machine learning to 1608 978-1-4673-1278-3/12/$31.00 ©2012 IEEE Proceedings of 2012 IEEE International Conference on Mechatronics and Automation August 5 - 8, Chengdu, China