Embodying and Improving SODA

We implemented variations on the Self-Organizing Distinctive State Abstraction model for robot state-discretization and task-learning in a Khepera robot. We conclude that hill-climbing as implementd in (5) is detrimental in embodied robots. Because the ultimate goal of developmental systems is intelligent, embodied robots, we believe that implementation in physical robots is an important test of robustness for proposed developmental mechanisms. In the fields of Adaptive and Developmental Robotics, a surprising number of systems produced through research do not involve physical robots. Instead, researchers focus on developing intelligence systems in virtual worlds, which are often simpler to use, more portable, and cheaper. Pfeifer and Scheier raise concerns that the traditional AI systems in symbolic worlds often lack robustness and generalization(4). Although simulated worlds have improved to more closely represent the noisy real world, they are still predictable and lacking in complexity. Even in a relatively controlled physical envi- ronment, a robot's motors often fail to produce the intended action and a robot's sensors often have to cope with statistically biased noise, neither of which is adequately addressed by existing simulations. If our ultimate goal is to produce robots and robot control mechanisms that can function independently in the real world, we need to test the robustness of our systems in physical robots as well as in simulation.