Sometimes active perception is required to find properties of objects that are not directly observable. For example, knowing the centre of mass of an object may be important for grasping, but it’s location may not be obvious from the object’s shape. However, some experiments may help to determine this and other hidden properties.To avoid performing unnecessary experiments, the robot’s visual system is used to create an internal model in a physics engine, where “thought experiments” are performed to determine which experiments in the real world will yield the highest information gain. This approach has been used to discover properties such as centre of mass, uneven friction in sets of wheels and to predict simple behaviours of other robots.

Publications include:

• O. Sushkov and C. Sammut. Active robot learning of object properties. IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012.
• O. Sushkov and C. Sammut. Feature segmentation for object recognition using robot manipulation. Australasian Conference on Robotics and Automation, 2011.