Practical Deep Learning Application in Robotics

The overall objective of this PhD is to further develop the art of deep machine learning algorithms in application to robotics in a way that is realistic and practical, utilising off-the-shelf embedded computing hardware typical for such applications. The task, to which the robot will be applied, is not yet fully decided, but will be novel, dynamic, challenging and including significant variation in the environment.

Very impressive progress has been made in the use of deep reinforcement learning applied to robotic manipulators. The robot learned control movements needed to grasp the objects and remove them from a tray. However, this work used up to 14 robots working in parallel and collected over 800,000 grasp attempts during the training phase. Thus, there is a high cost in terms of hardware (computing) and training time required that render current machine learning algorithms less than competitive in most robot applications.

The focus of the work will be to refine and optimise the application of deep learning methods to typical robotic applications. This will include investigation of different hardware platforms and optimisation of algorithms such as the utilisation of reduced floating-point machine precision for data representation and parallel processing to improve throughput.