Developmental learning for avoiding dynamic obstacles using attention

Dynamic environments, containing both static and moving objects, present a great challenge for collision avoidance. Inspired by early processing in the biological systems, the input to the learning system introduced here includes a range map and a motion map. Avoiding computationally expensive real-time models of the dynamic environment, the presented developmental learning method enables the system to learn to quickly identify relevant feature subspaces at every time instance from the real-time input sensory stream. An attention selection mechanism and the Incremental hierarchical discriminant regression (IHDR) were used to identify and switch to different feature subspaces at every time instance. The developmental learning method represents an architecture enabled trade-off among the limited computational power, the real-time requirements and the complex dynamic environments. Real-time simulations are implemented to provide promising results on the avoidance of moving obstacles.