University of Edinburgh

Data-driven machine learning techniques are popularly used in the field of robotics to inform autonomous decision-making and to perform control or path-planning. Supervised learning and reinforcement learning have been shown to be particularly amenable to canonical tasks that are integral to robotics applications. However, these techniques rely on data in the form of action-label (supervised), action-value (regression) or action-reward (RL) pairs, where the action is a path (or some other) execution by a real robot. e.g.

Curiosity guides humans to learn efficiently. It incentivizes us to spend more energy and time examining new, unexpected things, and to disregard those we fully understand already, to make our learning more efficient. Much of the vision learning that is done today is passive: learning systems are exposed to large amounts of training data, and learn from each sample multiple times, regardless of their current ability to recognize them at the time. This makes the process slow, specially given the increasingly large number of samples on datasets.

Human action recognition is a fundamental problem that underlies many applications in robotics, including interaction, home care, collaboration, etc. The actions that can be recognized by robots or computers today are often coarse and simplistic, in the sense that they are very different from each other; for example eating vs playing piano, or sitting vs standing. Both datasets and technology tend to be broad and crude. As human-robot-interaction becomes more natural we require more sophisticated technology for perceiving humans.

Vision is a key ability to humans as well as to robots to understand and extract information from real world and is crucial to safely interact with our environments. In contrast to human perception, state-of-the-art machine vision methods require millions of images and their manual labels to learn each visual task. This project focuses on designing machine learning and computer vision techniques that can help robots to learn multiple tasks from limited labelled data.

There are two main directions for potential projects:

How can we ensure that robots behave in a way as desired by humans? To ensure that the execution of complex actions leads to the desired behavior of the robot, one needs to specify the required properties in a formal way, and then verify that these requirements are met by any execution of the program.

The aim of this project is to develop a stringent approach to automatic programming of control systems based on an intermediate implicitly learned representation of the control task by a neural network. While deep neural networks (DNN) have been shown to be capable of solving control problems effectively based on learning from demonstration and reinforcement learning, the resulting representations are computational complex and lack immediate inspectability.

Responsible AI, and ethical decision making are major concerns in AI. While definitions and ideas differ on how to capture these notions, it is widely agreed that causality will play a major role. In our ongoing work, e.g. Responsible AI, and ethical decision making are major concerns in AI. While definitions and ideas differ on how to capture these notions, it is widely agreed that causality will play a major role. In ongoing work, e.g.

Multi-agent learning is an approach to solving sequential interactive decision problems, in which multiple autonomous agents learn through repeated interaction how to solve problems together. This includes agents working in a team to collaboratively accomplish tasks, as well as agents in competitive scenarios with conflicting goals. Reinforcement learning has emerged as one of the principal methodologies used in multi-agent learning, and a recent tutorial by Albrecht and Stone provides a basic introduction [1].

The issue of explanations for AI systems cooperating with humans has been a topic of considerable interest of late. But it is widely argued that current solutions that are based on local representations do not fully capture the reasoning behind the underlying decision.

The design of autonomous agents which can complete tasks in complex dynamic environments is a core area of research of modern artificial intelligence. A crucial requirement in such agents is the ability to interact competently with other agents (including humans) whose behaviours, beliefs, plans, and goals may be unknown. To interact with such agents requires the ability to reason about their unknown behaviours based on their observed actions, the context in which these actions took place, and other available information.