EPSRC Centre for Doctoral Training in Autonomous Intelligent Machines and Systems (original) (raw)

Autonomous systems in the real world may need to identify and interpret complex scenes, from moving vehicles to human activity. For example, robotic systems require such capabilities so that they can navigate in unknown environments, and augmented reality systems require methods for scene perception and object identification. AIMS will offer training in visual geometry, deep learning, perception, generative models, visual tracking, segmentation, path planning, visual navigation, robot learning, locomotion and motion planning. Domains of focus include urban & off-road driving, hazardous environments, smart infrastructure, space robots and healthcare.

AIMS recognises the increasing centrality of machine learning to autonomous systems. The data-driven approach offered by modern machine learning models has complemented the growth of datasets and computing power (particularly GPUs), which has resulted in machine learning transforming fields such as computer vision and robotics in the years since AIMS’s foundation in 2014. As such, AIMS now places Machine Learning as the hub of its training programme. Within machine learning, AIMS offers courses covering estimation & inference, signal processing, supervised & unsupervised learning, learning theory, Bayesian non-parametrics, graphical models, ensemble learning, neural networks and back-propagation, deep learning (representative models will include AlexNet, VGG-VD, ResNet, DenseNet), machine learning in production, optimisation, automatic differentiation, reinforcement learning, autonomous systems safety and governance, and much more. Focal applications are found in medical, smart city and space data domains.

Autonomous systems must be safe and dependable. For example, how do we ensure that the embedded software controller of the self- driving car does not crash, or that the implantable blood glucose monitor correctly identifies an abnormal range and raises an alarm? Verification via model checking provides automated methods to establish that given requirements are satisfied, but is challenged by the need to consider the complex interplay of discrete, continuous and probabilistic dynamics. Training will be given in modern control, deep reinforcement learning for process control, systems verification, adversarial inputs, symbolic controller abstraction techniques, computational game theory and multi-agent planning. Focal domains will include finance, factory automation, critical infrastructure and healthcare.

Many applications demand the seamless connection of intelligent devices in order to offer sensing, monitoring and actuating capabilities: these will form cyber-physical systems. Currently, cyber-physical systems face key technical barriers, including:

1. providing context-awareness (e.g. location), problematic in indoor environments;
2. overcoming the unreliability of sensors and actuators, often lacking calibration, quality estimation, energy management and fault detection capabilities;
3. Improving security and privacy concerns, both in peer-to-peer ad-hoc networks and cellular networks.

AIMS offers training in sensor and actuator networks, topology control, cross-modality learning, privacy and security, lightweight authentication and key management, wireless network security, anti- jamming/jamming resistance, passive/active threat models, verification of security systems, differential privacy, and hands-on embedded systems programming. Key applications include the smart city, critical infrastructure and sensitive healthcare data.

.