Minimalistic Vision-based Cognitive SLAM

The interest in cognitive robotics is still increasing, a major goal being to create a system which can adapt to dynamic environments and which can learn from its own experiences. We present a new cognitive SLAM architecture, but one which is minimalistic in terms of sensors and memory. It employs only one camera with pan and tilt control and three memories, without additional sensors nor any odometry. Short-term memory is an egocentric map which holds information at close range at the actual robot position. Long-term memory is used for mapping the environment and registration of encountered objects. Object memory holds features of learned objects which are used as navigation landmarks and task targets. Saliency maps are used to sequentially focus important areas for object and obstacle detection, but also for selecting directions of movements. Reinforcement learning is used to consolidate or enfeeble environmental information in long-term memory. The system is able to achieve complex tasks by executing sequences of visuomotor actions, decisions being taken by goal-detection and goal-completion tasks. Experimental results show that the system is capable of executing tasks like localizing specific objects while building a map, after which it manages to return to the start position even when new obstacles have appeared.