Deep Reinforcement learning of control and manipulation tasks using a tactile robotic hand.

点击翻译按钮获取中文摘要

In my proposed research I will explore the application of state of the art, supervised and unsupervised machine learning methods in the task of control and manipulation of a tactile robotic hand. Emphasis will be put on utilising deep reinforcement learning methods, under the co-supervision of area expert Dr Raia Hadsell of Google DeepMind.Deep reinforcement learning is a relatively unexplored area, particularly in its application to tactile sensing or robotics in general. It can be described as bridging the divide between high-dimensional sensory inputs and actions, which perfectly fits the proposed tasks of control and manipulation with tactile sensing. High dimensional sensory inputs are provided through images capturing tactile data, whilst actions can be a variety of control or manipulation tasks i.e. the movement of joints in a robotic hand.For the initial phase of the PhD the focus will be put on gaining a better understanding of how deep learning methods apply to a wider range of tasks based around a tactile sensor. This will lead into applying the methods to a full, sensorized tactile robotic hand, which will benefit from other work in the lab. The next phase will be to extend into unsupervised methods where deep reinforcement learning shows the most promise; groundwork for this methodology has been presented but this has yet to be realised with the TacTip or similar optical tactile sensing technology.