| Modular Robots: |
| Modular robots, or Modular Self-reconfigurable Robots (MSRs), are robots composed of many identical and interchangeable modules. The idea behind such a system is that a robot can rearrange these modules in order to change its own body shape or configuration. This ability to dynamically reconfigure would give the robot an increased robustness and generality. A MSR could change its shape to one suited for the task or replace broken modules with identical working modules. One way to view such a system is as mirroring a the cellular structure of living things. The modular robotics work I have been a part of is concerned with finding control strategies for the modules for such behaviors as global reconfiguration, locomotion, and object manipulation. This work has mostly centered on distributed control strategies, where each module runs an identical copy of predefined control software. As the number of modules scale up and the module size scales down, global control and planning for the modules may become infeasible. Further, because communication between modules may be limited, control strategies that use only localized information could prove very useful. This leads to the creation of control strategies that produce emergent behaviors. Each module acts on its own and the behavior of the entire MSR emerges from the local behaviors. Unfortunately, the creation of robust and useful distributed emergent behaviors is a very difficult task. My work on these problems has been as part of research teams at Xerox PARC and FXPAL. Specifically, at PARC I examined developing hand-coded control strategies for distributed control. Target behaviors included reconfiguration, locomotion, and object manipulation. These algorithms were tested in simulation. The development of such solutions turned out to be a highly difficult problem. The modules continued to find “new and creative” ways of breaking the hand-coded strategies. The following summer, at FXPAL, I worked on applying machine learning and genetic programming ideas to automatically generate decentralized behavioral control. (Publications) |
