EU scientists prove robots can learn to 'think'
EU-funded scientists have tested a groundbreaking theory that sees robots learning to 'think' about the actions they can perform on an object. The upshot is that robots can teach themselves by learning from their observations and experiences. This latest development is an outcome of the PACO-PLUS ('Perception, action and cognition through learning of object-action complexes') project, funded under the 'Information society technologies' (IST) Thematic area of the EU's Sixth Framework Programme (FP6) to the tune of EUR 6.9 million.
The PACO-PLUS project partners sought to test the so-called 'object-action complexes' (OACs) theory. OACs are units of 'thinking-by-doing' and this approach designs software and hardware that allows a robot to think about objects in terms of the actions that can be performed. For example, if a robot sees an object with a handle, the robot could grasp it. If it has an opening, the robot can potentially fit something into the opening or fill it with liquid. If it has a lid or a door, the robot can potentially open it. Objects therefore gain their significance by the range of possible actions a robot can execute upon them.
This opens up a much more interesting way for robots to think autonomously, because it fosters the possibility of emergent behaviour, complex behaviours which arise spontaneously because of quite simple rules, according to the partners.
The team's approach in many ways imitated the learning processes of young infants. As they encounter a new object, infants will try to grasp it, eat it, or bang it against something else, and as they slowly learn from trial and error that, for example, a round peg will fit into a round hole, their range of actions expands. A child's understanding also improves from watching other people.
PACO-PLUS conducted most of its work with Humanoid robots, robots shaped like people. 'Humanoid robots are artificial embodiments with complex and rich perceptual and motor capabilities, which make them… the most suitable experimental platform to study cognition and cognitive information-processing,' explained Dr Tamim Asfour from the Humanoids Research Group at the Institute for Anthropomatics at Karlsruhe Institute of Technology (KIT) in Germany, and co-coordinator of the PACO-PLUS project.
He said the group's work followed on the work carried out by Rodney Brooks, a leading robotics professor now based at the Massachusetts Institute of Technology (MIT) in the US. Dr Asfour explained that Professor Brooks 'was the first to explicitly state that cognition is a function of our perceptions and our ability to interact with our environment. In other words, cognition arises from our embodied and situated presence in the environment'.
Professor Brooks believed that moving and interacting with the environment were the difficult problems in biological evolution, and that once a species achieved that, it was relatively easy to 'evolve' the high-level symbolic reasoning of abstract thought, according to Dr Asfour. The reverse approach is taken by 'artificial intelligence', which believes if you develop enough intelligence, machine thought will be able to perceive and solve problems, he added.
The jury is still out on who is right, and the researchers admitted that while progress has been made, there is still no genuine intelligent robot candidates on the scene. 'That Hollywood interpretation is still a ways off, but the applications and demonstrators built by PACO-PLUS show that we are now, perhaps, on the right track,' the team said.