The 87th Installment
Modeling Collective Behavior through RoboCup
by Norifumi Watanabe,
Assistant Professor, Master Program of Information Systems Architecture
As the phrase “Two heads are better than one” says, there are many instances where something cannot be done by one person. However, a solution might be possible with a group. Many robots today implement pattern recognition utilizing high-precision cameras and sensors, and control functions that enable complex movements. Some robots are now at a stage where they can surpass human functions. However, there are still situations where an individual robot’s ability is limited. In such a situation, it will be necessary for humans and robots, and robots themselves to cooperate to realize capabilities that cannot be achieved by a single robot. Currently, we are conducting research to model human collective behavior and verify it by computer simulation in order to develop a robot that can achieve such collective behavior.
The term collective behavior can apply to various situations. Among these, we are focusing on a team sport, soccer, as the subject of the research. As an example of the conventional study of collective behavior, there was a study of evacuation behavior in emergencies. While it had an advantage of having a rather simple context, evacuation, which made it easy to regularize individual behavior, the situation was extraordinary and made it difficult to collect data. On the other hand, in the case of soccer, the movement trajectories and play information of players from all J.League games are currently accumulated, and the data is collected daily. Also, the rules of the game as well as the rules of the actions are clear and easy to model. However, not everything is coming up roses. Compared to evacuation behavior, the situation is more complex and difficult because there are multiple individual roles. Computer simulation is one of the methods which can be used to understand such a complex situation by modeling and can clarify the mechanisms within. In the case of soccer, the “RoboCup soccer simulation” exists as a simulator that models human movement.
RoboCup is an international robotics competition in which researchers and students of robotics and artificial intelligence gather and conduct research and development in order to “develop a robot soccer team which beats the human world champion team by 2050.” RoboCup Soccer Simulation Leagues started in 1996. In the early stages of RoboCup, matches of 11 vs 11 players (agents) on computers would take place. Currently, we have the 2D League, which uses a 2D map, and 3D League, which uses a 3D field. Basic actions such as a kick, dash, and tackle are implemented on individual agents, and more actions can be added depending on the position, such as forward and defense. In order to make the actions as close as possible to those of a human, parameters such as speed, power, and stamina are set, and agents will be limited by those parameters after performing certain actions. Unlike soccer games released for consoles, these agents cannot be controlled from outside. Based on the information gained from the line of sight such as player agents and position and direction of the ball, as well as based on the instructions from the coach agent, each player agent autonomously plays the game using their own decision-making.
The following three points are important when creating the RoboCup agents.
1. Complete distributed multi-agent system
2. Incomplete information processing
3. Real-time processing
The distributed multi-agent system (1) is a research field in which autonomously acting agents interact to solve problems. In RoboCup, each agent has to decide their own actions based on their own abilities and positions of the ally players. The incomplete information processing (2) is important because in RoboCup, the agents have their own line of sight, which means they can only see a part of the field, so they have to decide their actions by speculating the overall movement from the information they can gain. The real-time processing (3) is important because the simulator allows the match to proceed without waiting for the decision of agents, so if the decision-making is taking time, there will be instances where the situation changes but the actions are not taken in time.
So what we’re working on right now is research which models the common intention among the players, which is in essence what is considered to be constantly happening with human soccer players during a game, and implement it to agents in RoboCup. By figuring out the common intention to realize team play among players, agents speculate what team play actions to perform from the positions of the other agents and ball, and decide their own actions within the situation. We are now extracting the data of how the players speculate in different situations from J.League matches, and modeling the common intention. By implementing further structured models on RoboCup agents and replicating those as collective behavior in the simulator, we assess if it is appropriate as a human algorithm, and if not, we assess how it is wrong in order to better understand human collective behavior. After creating a RoboCup team based on these research approaches, we participated in the international RoboCup competitions in 2016 and 2017, and in 2017, we reached the rank of 8th place worldwide. In the future, we will model other types of collective behavior, not just soccer, and aim to develop robots that can collaborate with humans and other robots in complex situations that involves drastic changes.