AbstractsPhysics

Reaching to grasp is a fundamental human action. In this thesis, I present a new computational platform for simulating reach-to-grasp actions and exploring how they are learned and performed. There are three innovations in the platform. Firstly, many existing platforms for simulating reach-to-grasp actions hard code a definition of 'a stable grasp'. My aim was to create a platform based on an existing physics engine (the JMonkey game engine), so that the stable grasp is represented using general-purpose definitions of force and friction. Secondly, the platform implements a new model of the soft finger pads of the human hand, and of the finger's tactile mechanoreceptors. Each finger pad is modelled as a lattice of solid objects connected by springs to create a deformable surface. This allows a good simulation of the mechanoreceptors which respond to local finger pad deformations, and to other kinds of tactile stimulus. Finally, the platform supports a novel model of the motor controller responsible for generating reach-to-grasp actions. Most existing computational models assume that the hand's trajectory to the target is precomputed in advance, but there is evidence that this does not happen in the primate reach/grasp neural pathway. The motor controller in my system is a combination of a low-level feedback controller which tries to move the hand and arm into a learned goal state, and a high-level controller which perturbs or moves this goal state to create a virtual target location for the low level controller to reach towards. Combining these controllers allows for complex actions to be learned, without precomputing trajectories.