Place: Room 621/622, GP South (Building 78)
Time: Thursday 17th August, 10:30 morning Tea, 11:00am seminar
Speaker - Dr Scott Bolland
Key Centre for Human Factors as Applied Cognitive Psychology The University of Queensland Thursday
Title: "How to self-teach a young robotic dog new tricks: an exploration in developmental robotics"
Although human high-level cognitive processing has often been described in the literature as "symbolic" in nature, flexible real-world reasoning and problem-solving are often sensitive to subsymbolic task and object related features. For example, consider the task of changing a light bulb and the reasoning processes involved in selecting an object on which to stand in order to gain the appropriate height. Although the resulting selection can be expressed symbolically (e.g., "the black chair"), the reasoning process itself is sensitive to low-level task specific features such as the stability, shape, and weight-bearing characteristics of the available options. In developing artificially intelligent "thinking systems", it is doubtful that such subsymbolic sensitivities can be explicitly programmed.
Instead, learning appropriate grounded representations through interacting with the world is an important (and perhaps necessary) characteristic of artificially intelligent embodied systems.
In learning to interact with the world, intrinsic reward systems have proven useful in training robots to develop task independent competencies and appropriate internal representations without the need for explicit teaching.
For example, studies in developmental robotics have shown that the use of generic learning schemes (such as reinforcing actions that lead to predictable changes in the environment) can be used to teach a robot about affordances (such as the fact that only objects at a certain distance can be grabbed, or that only objects that communicate back are worth talking to).
However, many of the recently proposed algorithms for self-directed learning have only been demonstrated to work in limited hand-crafted domains, not scaling well to more natural environments. This talk will highlight some of the main features of these algorithms that restrict their utility, and provide biologically motivated architectural and functional extensions (based on what is known about the role of the hippocampus and the dopamine system on attention and learning) that substantially improve their performance. Experiments will be demonstrated on a Sony AIBO ERS-7 robotic dog.