Online learning for dynamic robotic manipulation tasks

Present approaches to tactile sensing and control require large amounts of data to train machine learning algorithms, or other statistical methods, to transform low-level sensory data into high-level information such as contact position, angle and force. Once a suitable model is learnt from data it is then used to within a control policy to complete the desired task. While this approach is effective, it is far from efficient. Moreover, the overall performance is limited by the initial training data; performance will not improve as more data is collected during the task and novel sensory inputs are likely to prevent completion of the task.

Instead, by defining a high-level objective function it is possible to remove the initial training step and, for example, track contours and perform other tasks. This objective function encodes information about what the task is without prescribing how it is performed. For example, a simple contour following task could be constructed by generating a single reference point with the tactile sensor in contact with the contour. The objective function is then the (suitably defined) distance between a measurement point and the reference point. Moving the sensor such that the objective function is minimised, combined with an appropriate exploration policy means that the contour will be tracked.

This approach, which by-passes the construction of a sensor model, removes the limitations surrounding the presence of novel sensory inputs. In all cases, the control policy simply seeks to iteratively minimise the objective function (e.g., gradient descent).To improve matters further, it is possible to combine this objective function with online machine learning; the iteration is accelerated by learning a local interaction model from the accumulated sensor readings and the known movements between them. Hence from a single reference point, a robot manipulator can learn how to interact with its surroundings as it is carrying out useful tasks. As such, this approach has the potential to dramatically change the efficiency and effectiveness of tactile sensing and control.

This project seeks to generalise the capabilities ofexisting tactile manipulation strategies to situations where novel sensoryinputs are present. As such, there are a wealth of possible avenues ofresearch, ranging from improving the learning algorithms (e.g., how to dealwith uncertainty in the inputs, and how to deal with non-stationary situations)through to how to generalise the control policies to a wide range of differentmanipulation tasks (e.g., grasping and general manipulation).

To apply for this PhD project contact David Barton ( with your CV.