Natural cost functions for contact point selection in grasping

Paul Schrater
University of Minnesota

When reaching to touch or lift an object, how are contact points visually selected? In this talk I will formulate the issue as a statistical decision theory problem that requires minimizing the expectation over a suitable loss function. However, it is the nature of this loss function that is the heart of the presentation. In the first part of the talk, I will show how contact points for two fingered grasp can be optimally chosen, given a plan for the grasped object’s motion. The basic assumption is that the minimum control framework used to predict hand trajectories should also apply to the control of the grasped object. The cost function on the object’s motion can then be rewritten in terms of finger placement and contact, inducing a cost function on finger contact points. I will present human reaching data that supports this idea. In the second part of the talk, I will present evidence for a decomposition of the natural cost function for reaching into task completion and motor control components. The issue can be framed as follows: In many reaching tasks there are a set of contact points that are equivalent in terms of task completion cost—touching a line, for example. In generating a path, the ambiguity is broken by motor control cost, which distinguishes the minimum control point of the set (e.g. the closest point on the line). This unique target point could be selected to generate a simple feedback control strategy of minimizing distance to the target. Alternatively, a feedback control strategy could be based directly on a lumped cost function. These two strategies behave differently under a perturbing force field mid-reach: the first corrects the perturbations, while the second “goes with the flow” to contact the new minimum control point within the task completion set. I will present data supports the idea that reaches "go with the flow", adapting to external perturbations. This suggests that the brain visually encodes and adaptively uses the set of viable contact points. Finally, I will discuss why the contact point selection problem is important for understanding the sensory demands made by the motor control system during reaching.

Presentation (PowerPoint File)

Back to Probabilistic Models of Cognition: The Mathematics of Mind