Soccer Behavior Design for a Quadruped Robot
(Prof.
T. ARAI and Prof. H. YUASA)
RoboCup (The Robot World Cup Soccer Games) is a standard problem of Artificial Intelligence. The team ARAIBO has participated in SONY Four-Legged Robot League of RoboCup since 1999.
Soccer robots are required to decide its action with uncertainly information, since the robot does not have enough time to obtain information. For example, the robot must decide properly whether it walks or observes for self-localization since walking actions fuzzify the robot’s position information, and an observation action for self-localization consumes time. This indicates that the robot’s position information always contains a measure of uncertainty.
We proposed a real-time decision making algorithm based on Dynamic Programming (DP), and applied to solve a problem that the robot approaches the ball quickly from proper direction. We added one axis which denotes uncertainty of self-localization results to the state space, besides the robot’s pose (3-D) and the ball’s position (2-D) (Fig.1). Then we executed DP on PC on the assumption that time consumption of each action is a negative reward, and obtained State-Action Map, which is a database of the optimal action at each state.
We compressed this six-dimensional huge map with Vector Quantization (VQ) so as to implement on the robot. Using the differences of the values between the optimal action and the others as distortion measure of VQ minimizes the total loss of optimality (Fig.2).
Fig.3 shows an example of the robot’s behavior based on the compressed map. The robot approached the ball from proper direction with only one observation action. As this example, we validated that the map restrained unprofitable observations while the robot was able to carry out the task, even if the position information of the robot was very uncertain.
Keywords: RoboCup,
Quadruped Robots, Soccer Behavior
References
Takeshi Fukase, Yuichi Kobayashi, Ryuichi Ueda,
Takanobu Kawabe, Tamio Arai: “Real-time Decision Making under Uncertainty of
Self-Localization Results,” The RoboCup 2002 International Symposium.
Fig.1State transition
in the extended state space
Ryuichi Ueda, Takeshi Fukase, Yuichi Kobayashi, Tamio Arai, Hideo Yuasa
and Jun Ota: “Uniform Monte Carlo Localization - Fast and Robust
Self-localization Method for Mobile Robots,” Proc. of ICRA-2002.
Fig.2 The difference of action-value
Fig.3 An example of experiments