Learning to Map Degrees of Freedom for Assistive User Control - Towards an Adaptive DoF-Mapping Control for Assistive Robots.

This paper presents a novel approach to shared control for an assistive robot by adaptively mapping the degrees of freedom (DoFs) for the user to control with a low-dimensional input device. For this, a convolutional neural network interprets camera data of the current situation and outputs a probabilistic description of possible robot motion the user might command. Applying a novel representation of control modes, the network’s output is used to generate individual degrees of freedom of robot motion to be controlled by single DoF of the user’s input device. These DoFs are not necessarily equal to the cardinal DoFs of the robot but are instead superimpositions of those, thus allowing motions like diagonal directions or orbiting around a point. This enables the user to perform robot motions previously impossible with such a low-dimensional input device. The shared control is implemented for a proof-of-concept 2D simulation and evaluated with an initial user study by comparing it to a standard control approach. The results show a functional control which is both subjectively and objectively significantly faster, but subjectively more complex.