In recent years, robots are needed to work in more complex, various and dynamic environment, and flexibility, extendability and fault-tolerance of systems that make the system adaptable for changes of complex environment and various task demands are demanded. In previous works, optimal motion planning method is based on the explicit models of robots and environments. However, this kind of approach is difficult to adopt in the above-mentioned situation. This is because it is difficult to prescribe the appropriate situations in which the robot will stand and the behaviors appropriate for them, and hence it becomes impossible to design appropriate relationship between input and output. In order to deal with this kind of situations, robots should be designed as a emergent system, which has abilities to change own internal structure through interactions with complex and dynamic environment, to emerge novel structure, to self-organize or to evolve itself.
Regarding this background, our research group aims to realize robots as emergent systems. Concrete research issues are following:
1) Reinforcement learning in partially observable Markov decision process with autonomous state segmentation
Reinforcement learning is widely noticed as a promising method for behavior acquisition of robots. However, in previous works, applications of reinforcement learning is limited to robots that works in real world because of the following two limitations: (1) The environment is modeled as Markov decision process, (2) State-space should be descretized by designers in advance. In this research, we model the environment as partially observable Markov decision process and autonomously segment the space of observation that robot observes according to the interaction with environment. Using this method, we aim to propose a reinforcement learning system to solve the above-mentioned problems.
2) Autonomous construction of state-space using visual information
Robots working in real world need to be adaptable for various I/O structures. In our research, we focus on visual input as multi-dimensional input and deal with a process to autonomously extract useful information from visual input according to value signal without any prior knowledge about the world. Tactile sensor information is used as the value signal. We study the method to self-organize state and action appropriately.
3) A principle of cooperation of multiple dynamic components and generation of a global order
We are studying a fundamental and theoretical research about a principle where multiple distributed dynamic components behave cooperatively and total system behaves meaningfully.
Applying the result of our research, a super-parallel image processing system using dynamic image pixels can be realized by regarding dynamic components as pixels on the image. On the other hand, if we consider dynamic components as autonomous mobile robots, we can establish a principle to realize cooperative behavior of multiple robot system. Thus, our research focuses on a very general cooperation principle which is applicable for various objectives.