Motion Synthesis, Learning and Abstraction through Parameterized Smooth Map from Sensors to Behaviors

Abstract

The integration theory of reactive behaviors is to be discussed in this paper. A linear emerging model is adopted where the motion of a robot is represented as the weighted linear sum of reactive behaviors. The weights are defined as differentiable nonlinear functions of sensor signals and parameters. The functions can represent logical if-then rules as their extreme cases. The sensor space model is introduced to relate the sensors and the behaviors and to determine the parameters. We establish a learning method based on the sensor space model, where the parameters are systematically tuned through iteration of trials such that the sensor signals converge to the given teacher signals. A nonlinear dynamics in the sensor space model is also proposed to allow fluctuation for the future global search. The learning method is applied to the reactive grasp of a three-fingered robot hand. We integrate 48 kinds of sensor signals and 29 primitive behaviors. The experiments indicate that the emerging model allows us to use the semantics to initially program the nonlinear functions for the weights. The learning experiments successfully illustrate the usefulness of the proposed learning method.