"Connecting experiences": Towards a biologically inspired memory for developmental robots

Ajaz Ahmad Bhat, Vishwanathan Mohan, Francesco Rea, Giulio Sandini, Pietro G. Morasso

Abstract: We all inhabit a continuously changing world where practically neither everything can be known nor can everything be experienced. In order to ultimately survive, diverse `context relevant' chunks of knowledge emerging from one's multiple past experiences have to be `connected' and exploited flexibly in the context of the present to ensure smooth realization of goals. Understanding how the brain achieves such creative and generative use of experience is a central challenge facing both neuroscience and cognitive robotics today. This article is an exploration in this direction through playful, cumulative experiences of humanoid iCub. A biologically inspired neural framework for organizing, recalling and flexibly combining memories of multiple experiences of the robot under novel situations is proposed. The architecture is validated through both a simple staking task and then a real world industrial “fuse box” assembly task where recall of context specific past experiences trigger simulation of consequent future events in time hence allowing the robot to realize an otherwise unrealizable user goal: Pushing an object that is non-existent in the scene (but could exist if the robot creates it by exploiting its past experiences). Through the resulting behaviors, we emphasize that neural mechanisms for organization of episodic memory is a fundamental design feature necessary for the survival of autonomous robots in a real world `where neither everything can be known nor can everything be experienced'.