Spatial Reasoning and Planning with Deep Neural Networks

Lead Research Organisation: University of Oxford
Department Name: Autonom Intelligent Machines & Syst CDT

Abstract

Context & Potential Impact - Performing tasks autonomously in novel environments is a major challenge for many applications of robotics, from indoor robots to autonomous vehicles. In classical robotics, such tasks have been separated into smaller components: localisation, mapping, planning and navigation. Techniques such as Simultaneous Localisation and Mapping (SLAM) have been popular for localisation and mapping, while optimisation, graph search and control theory has been prominent in the field of planning and navigation. While these methods can solve well-specified individual problems, their performances are upper-bounded by the assumptions made by each particular component.
Advances in neural network architectures, together with successful integration of these models into Reinforcement Learning (RL) frameworks, have transformed the field of computer vision and robotics, allowing a learning-based approach to problems traditionally solved by manually implementing insights of domain experts. Learning-based approaches have two major advantages. Firstly, learning-based algorithms can keep improving and adapting to the application domain with more availability of data, whereas manually implemented methods are fixed and do not learn to adapt. Secondly, learning-based methods are capable of automatically discovering inherent regularities and characteristics of the application domain and exploiting them to improve their performances without having such strategies hardcoded.
Aims & Objectives - In this DPhil research project, we develop end-to-end trainable systems for spatial reasoning and planning, where components for localisation, mapping, and planning are purposefully engineered. We hypothesise that having explicit constraints on the capability of the components will allow the system to learn transformations and policies that could generalise to unseen scenarios, and could be easily fine-tuned for new environments and tasks.
Research Method Novelty - The problem of learning-driven autonomous navigation is an open-ended problem, and there isn't yet a system that takes a fully learning-based approach without compromising robustness and interpretability. Many previous works either assume a fully known environment, or use neural network architectures that do not have a guarantee of learning robust internal representations of spatial memory usable for general purposes. Therefore, we focus on integrating a family of memory networks into our system, which should give a more interpretable representation of spatial memory. Having a better state representation would also unload the burden of learning a complex policy at the RL phase. We will also develop an end-to-end differentiable planning algorithm that can work with these intermediate representations so that better representations are learnt with more training data.
Alignment to EPSRC's strategies and research areas - Artificial intelligence technologies, Image and vision computing, Information systems, Robotics and Software engineering

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S024050/1 01/10/2019 31/03/2028
2242819 Studentship EP/S024050/1 01/10/2019 30/09/2023 Shu Ishida