Autonomous Environment and Task Adaptation for Robotic Agents

This paper investigates the problem of improving the performance of general state-of-the-art robot control systems by autonomously adapting them to specific tasks and environments. We propose model- and test-based transformational learning (MTTL) as a computational model for performing this task. MTTL uses abstract models of control systems and environments in order to propose promising adaptations. To account for model deficiencies refrom abstraction, hypotheses are statistically tested based on experimentation in the physical world. We describe XFRMLEARN, an implementation of MTTL, and apply it to the problem of indoor navigation. We present experiments in which XFRMLEARN improves the navigation performance of a stateof-the-art high-speed navigation system for a given set of navigation tasks by up to 44 percent.

• Type: Conference
• Year: 2000
• Where: ECAI
• Authors: Michael Beetz, Thorsten Belker
• Keywords: Artificial Intelligence, Control Systems, ECAI 2000, State-of-the-art Robot Control, Test-based Transformational Learning