Robot Programming through a Combination of Manual Training and System Identification

In this paper we present a novel procedure to obtain the control code for a mobile robot executing sensor-motor tasks. The process works in two stages: First, the robot is controlled by a human operator who manually guides the robot through the sensor-motor task. The robot’s motion is then “identified”, using the NARMAX system identification technique. This process, which we refer to as RobotMODIC (robot modelling, identification and characterisation) has distinct advantages over existing robot programming techniques: 1. The RobotMODIC process allows very rapid code development without the need of iterative refinement, 2. it requires little a priori knowledge of the task performed — the robot is merely driven manually during the training phase, 3. the RobotMODIC process produces extremely compact code, 4. the generated coupling between perception and action can be analysed, using mathematical methods, and 5. the generated code is largely robot-platform independent. The RobotMODIC process represents a step towards a science of mobile robotics, because it reveals fundamental properties of the sensor-motor couplings underlying the robot’s behaviour through its transparent and analysable modelling method.