Learning Task-specific Memory Policies

S. Rajendran and M. Huber (USA)


Focus of Attention, Memory Policies, Robots, AI agents


Effective AI agents and robots require the ability to adapt to real world situations and perform multiple tasks. This requires them to take into account the important sensory information. Extraction of this information can be made tractable using mechanisms of focus of attention that select perceptual features that have to be processed. This mechanism alone however is inadequate for tasks in real world situations since it still requires the robot to maintain all past information, rendering decision making computationally intractable. This requires the robots and AI agents to have the capability to remember only the past events that are required for successful completion of a task. Here we present an approach (illustrated using block stacking and block copying tasks) that extends a previous focus of attention mechanism by incorporating short term memory to remember past events. The result is a task specific control, sensing, and memory policy.

Important Links:

Go Back