I am a final year PhD student researching with the Applied Artificial Intelligence Lab (A2I) attached to the Oxford Robotics Institute, University of Oxford. My research focuses on the application of AI and machine learning to better enable robots to robustly and effectively operate in complex real-world environments. In my PhD I am attempting to better harness the potential of radar for autonomous driving and robotic applications, through the application of state-of-the art machine learning approaches. For more information see my publications below. More recently I have pivoted more towards machine vision - I have spent the last three months interning with Nvidia developing systems for large scale semantic 3D mapping and scene understanding using Stereo cameras for AV applications.
I am currently looking for jobs at the intersection of robotics, autonomous driving and machine learning.
Here you can find a complete list of my publications. For more info (including videos!) press the drop down arrow or follow the various links to more project content.
In 2022 International Conference on Robotics and Automation (ICRA) (Under Review)
The current state of the art in radar odometry, MbyM, provides robust and accurate odometry measurements through an exhaustive correlative search across discretised pose candidates. However, this dense search creates a significant computational bottleneck which hinders real-time performance when high-end GPUs are not available. Utilising the translational invariance of the Fourier Transform, in our approach, f-MbyM, we decouple the search for angle and translation. By maintaining end-to-end differentiability a neural network is used to mask scans and trained by supervising pose prediction directly. Training faster and with less memory, utilising a decoupled search allows f-MbyM to achieve significant run-time performance improvements on a CPU and to run in real-time on embedded devices, in stark contrast to MbyM. Throughout, our approach remains accurate and competitive with the best radar odometry variants available in the literature -- achieving an end-point drift of 2.01% in translation and 6.3deg/km on the Oxford Radar RobotCar Dataset
Rob Weston, Oiwi Parker Jones and Ingmar Posner. In 2021 International Conference on Robotics and Automation (ICRA)
Simulating realistic radar data has the potential to significantly accelerate the development of data-driven approaches to radar processing. However, it is fraught with difficulty due to the notoriously complex image formation process. Here we propose to learn a radar sensor model capable of synthesising faithful radar observations based on simulated elevation maps. In particular, we adopt an adversarial approach to learning a forward sensor model from unaligned radar examples. In addition, modelling the backward model encourages the output to remain aligned to the world state through a cyclical consistency criterion. The backward model is further constrained to predict elevation maps from real radar data that are grounded by partial measurements obtained from corresponding lidar scans. Both models are trained in a joint optimisation. We demonstrate the efficacy of our approach by evaluating a down-stream segmentation model trained purely on simulated data in a real-world deployment. This achieves performance within four percentage points of the same model trained entirely on real data.
Dan Barnes, Rob Weston, and Ingmar Posner. In 2020 Conference on Robot Learning (CoRL)
This paper presents an end-to-end radar odometry system which delivers robust, real-time pose estimates based on a learned embedding space free of sensing artefacts and distractor objects. The system deploys a fully differentiable, correlation-based radar matching approach. This provides the same level of interpretability as established scan-matching methods and allows for a principled derivation of uncertainty estimates. The system is trained in a (self-)supervised way using only previously obtained pose information as a training signal. Using 280km of urban driving data, we demonstrate that our approach outperforms the previous state-of-the-art in radar odometry by reducing errors by up 68% whilst running an order of magnitude faster.
Rob Weston, Sarah Cen, Paul Newman, and Ingmar Posner. In 2019 International Conference on Robotics and Automation (ICRA), pp. 5446-5452. IEEE, 2019
Radar presents a promising alternative to lidar and vision in autonomous vehicle applications, able to detect objects at long range under a variety of weather conditions. However, distinguishing between occupied and free space from raw radar power returns is challenging due to complex interactions between sensor noise and occlusion. To counter this we propose to learn an Inverse Sensor Model (ISM) converting a raw radar scan to a grid map of occupancy probabilities using a deep neural network. Our network is selfsupervised using partial occupancy labels generated by lidar, allowing a robot to learn about world occupancy from past experience without human supervision. We evaluate our approach on five hours of data recorded in a dynamic urban environment. By accounting for the scene context of each grid cell our model is able to successfully segment the world into occupied and free space, outperforming standard CFAR filtering approaches. Additionally by incorporating heteroscedastic uncertainty into our model formulation, we are able to quantify the variance in the uncertainty throughout the sensor observation. Through this mechanism we are able to successfully identify regions of space that are likely to be occluded.
Here you can find a selection of blog posts that I have written over the last few years:
Here you will find some supporting material for my most recent submission to ICRA 2022. More will be added over the next few months. Watch this space...
Here you will find a landing page supporting the project "There and Back Again: Learning to Simulate Radar For Real World Applications" presented at the International Conference on Robotics and Automation (ICRA) 2021.
During the first lockdown, with more free time on my hands, I began to think a little more philosophically about the foundations of mathematics. In particular how can we know for sure that maths is correct? I discovered that at the start of the 20th century this was a question that an enormous amount of time and effort was devoted to trying to answer. Of the solutions proposed, the work of Zermelo and Fraenkel alongside the axiom of choice, commonly abbreviated to ZFC, remains one of the most complete answers to date. In this post I begin by looking at why ZFC is necessary, what it is, how it might be used to ground many of the important foundations of mathematics today.
Since their conception in 2014, the use of generative adversarial networks has exploded throughout machine learning, vision and robotics. Alongside novel application domains, much time and effort has been devoted to developing new architectures and training approaches. Modern instantiations of the GAN training paradigm are responsible for some truly remarkable results. However, in trying to access the wealth of material that is out there, the sheer quantity can be difficult to penetrate. This blog is intended to be a brief introduction to generative adversarial networks and their development over the last few years. It is by no means exhaustive but simply communicates my ideas of what GANs are and my interpretation of how we got here. So without further a do...
Here you will find the slides I presented at several reading groups designed to be a general intro to Bayesian Inference and linear regression. Much of the content is based on the Murpy and Bishop machine learning books which I thoroughly recommend if you want to find out more. If you have any questions or want to discuss anything please drop me an email.
This page contains a cheat sheet covering much of the first year electrical engineering course at the University of Oxford. It grew out of a collection of notes I made for myself whilst teaching undergraduate tutorials over the last couple of years and I hope it will be a useful resource for others. We begin by considering Maxwells equations which - remarkably - are able to describe the entireity of classical electromagentism and optics and underpin the workings of all the electrical devices covered below. Next , linear circuit devices are introduced, including resistors , capacitors and inductors, and a selection of methods for analysing both AC and DC circuits are summarised. If you have any questions / spot any mistakes please do drop me an email.
Xmen is a python api and command line tool I developed from the ground up for writing and running experiments. The user writes clean python code with minimal modifications and xmen takes care of the rest, including configuring, running and recording experiments, interfacing with the slurm job scheduler and visualising experiments from the command line.
A lot of work was put into developing xmen. It is now a key component of my workflow. Please give it a go and let me know what you think! For more info please see the xmen project website
In building this website I developed my own jeckyll theme. This was a complete first for me and I learnt a lot through the process . The aim was to develop a theme that is research friendly, with built in math support, easy page navigation and support for project pages. Much of the appearance is inspired by latex. Some work still needs to be done to get the theme into a sharable state for an official release. In the meantime you can find the sourcecode for this site here.
Over the lockdown vim and vim plugins became a key part of my workflow. For anyone who is interested you can find my vim config here.