Robotics Research

Applications: Learning and Adaptation

In order to be successful, most robots need the ability to react intelligently to changes in their environment (such changing environments are often labelled dynamic). In order to react with intelligence, many robots are designed with the ability to actively adapt themselves to new conditions, or even learn new skills which could help them to operate better. It is this ability to learn and adapt which is generally associated with intelligence, and is an area of great interest to robotics researches around the world.

Traditionally, intelligence was often modelled as an expert system – a large database in which pieces of knowledge are stored, with an overall program to sort and use this information. However, while successful in some situations, this sort of system was not ideal because it required the existence of an outside ‘expert’ to provide the knowledge in the first place, and was inflexible and performed poorly when confronted with something new. Far more promising results have been seen more recently with the use of biologically inspired approaches similar to the kind often seen in control systems. Theses systems attempt to approximate the way in which living things learn and evolve.

Neural networks, genetic and evolutionary algorithms and other such constructs are able to adapt to changes in the environment in a robust manner, and if designed correctly can often produce optimal solutions to problems independent of direct outside assistance. Unlike so–called expert systems, many of these systems can be easily trained to perform a variety of tasks without the need for expert a priori knowledge. 

Learning and Adaptation Research in Australia

• The Robotics Laboratory at the University of Queensland has conducted a number of projects based on the learning and adaptation seen in biological systems. Such projects include:
  • CORGI, a vision guided robot that learned to perform vision based foraging tasks.
  • RatSLAM, a system that learns both the visual cues and the map of an environment using methods based on studies of the rodent hippocampus.
  • Learning the control of the RoboRoos soccer playing robots using learning models based on the cerebellum.
  • The use of genetic algorithms to learn gait and control parameters on the GuRoo humanoid robot.
• Alan Blair at UNSW is conducting research into co-evolutionary algorithms for a variety of applications, including robotics.
• Stephan Chalup and the Interdisciplinary Machine Learning Research Group at the University of Newcastle are working in such areas as machine learning, data mining, evolutionary computing and reinforcement learning.
• The IRRC are in the process of a developing a humanoid robot which is intended to have the capacity to learn tasks for manipulating the environment from human operators.
• Claude Sammut from UNSW is involved in research into machine learning, particularly in the areas of inductive logic, behavioural cloning, reinforcement learning and speaker identification.
• The Robotic Systems Laboratory at ANU us working on reinforcement learning in mobile robots, and have devised a system for applying this form of learning to systems with continuous variables.