Agents engaged in lifelong learning can benefit from the ability to acquire new concepts from continuous interaction with objects in their environments which is a ubiquitous ability in humans. This paper advocates the use of sensorimotor concepts that combine perceptual and actuation patterns. Related representations to sensorimotor concepts are Predictive State Representation in dynamical systems, Affordance Based Concepts in language and Skills in reinforcement learning. The paper proposes a system for learning generalized sensorimotor concepts from unsegmented interactions between the agent and the objects in its environment that works in continuous action and observation spaces and in the same time require no reinforcement signals. A proof-of-concept experiment with the proposed system on a simulated e-puck robot is reported to support the applicability of the proposed approach.
展开▼
