Mental simulation of actions for learning optimal poses

Abstract

Mental simulation of actions is a powerful tool for allowing cognitive agents to develop Prospection Capabilities that are crucial for learning and memorizing key aspects in challenging actions. In particular, this study focuses on the initial or final posture of actions and provides a computational tool that allows an agent to evaluate their feasibility and appropriateness. Such tool is a kinematic network, equivalent to an internal body schema, that allows a cognitive agent to generate simulation-states that reach the goal with a comfortable final posture, by exploiting the redundancy of the kinematic network. This is obtained by activating and integrating in the network dynamics three types of virtual force fields: 1) Focal force field applied to the end-effector, related to the goal of the action; 2) Range of Motion force fields, applied separately and independently to each degree of freedom in order to preserve the natural joint limits; 3) Postural force field, applied to the pelvis area, for maintaining the projection of the center of mass of the body model inside the support base. The efficacy of this approach is demonstrated in relation to a simple task: reaching a heavy load in order to lift it and then shifting it forward before dropping it on a table. The mental simulation model attempts to provide a kinematic template compatible with the overall plan and the postural/articular constraints, as a function of the initial position of the body relative to the load. The simulation may fail and this indicates that the chosen initial posture is inappropriate for the task. Successful simulations can also be evaluated in terms of precision and effort by monitoring the peak torque required of each joint actuator. Optimal or at least sub-optimal solutions can be memorized in episodic memory, thus accruing the know-how of the agent.