The Applied Research group at Five is a collection of Research Scientists and Research Engineers, spread across our Cambridge, Oxford and Edinburgh offices, dedicated to investigating the company’s key longer term problems. All Applied Research projects are targeted at being capable of providing significant enhancement to some aspect of our products. However, whereas the company’s Engineering teams are focussed on work that will have an impact over the period of a single sprint or perhaps a few months, the Applied Research work will typically take 3-12 months to have an impact.
The company’s mission of providing tools to other companies to help speed up the development and ensure the safety of their Autonomous Driving System (ADS) systems requires us to understand all the key problems of developing Autonomous Driving Systems. In addition our simulation products require some flavours of some of the components of an ADS within our product itself, for example offline versions of perception components and offline versions of prediction and planning components.
In order for an ADS to drive safely and make suitable progress in the presence of other road users it is necessary for it to understand the relevant aspects of the world around it and then make suitable decisions. Our focus is on the algorithms that allow an ADS to perceive the world around it using information from sensors such as cameras, RADAR, LIDAR, IMU, GPS etc, and on the algorithms that allow an ADS to make safe driving decisions in the presence of uncertain perception information. In several cases these algorithms use machine learning (ML) as part of their operation and an additional crucial focus is on how we can become confident of the performance and safety of algorithms within an ADS and of the whole ADS.
To address the perception problem we have activity in computer vision and machine learning including topics such as:
- Odometry (our vehicle’s movement) and localisation (its position on a map)
- Scene and object reconstruction, shape retrieval and completion
- Object detection, tracking and segmentation (in sequences of images and point clouds)
To improve accuracy and reliability of perception it is necessary to fuse information from these multiple sensors and detectors using an understanding of how likely to be correct the information is. Also, we believe it is impossible to label sufficient data to solve or verify or validate these real world problems, so it is also essential to find techniques able to make use of unlabelled data. Hence our interest in topics such as:
- Uncertainty, calibration, anomaly detection, robustness, adversarial attacks, domain shift
- Self-supervision, domain adaptation, few shot learning
To plan the vehicle’s path through the world requires prediction of the behaviour of other agents around the vehicle and an understanding of the risk of taking different actions. Hence our interest in topics such as:
- Agent prediction (including goal extraction in the context of given road situations)
- Risk aware path generation (in the presence of uncertainty)
- Control theory
Finally, to build a verification and validation platform for ADS requires flexible, realistic models in simulation, a clear understanding of Operational Design Domains (ODDs) and how to cover all situations related to that ODD in simulation. Hence our interest in topics such as:
- Surrogate perception models, realistic agent models, and generative realistic HD maps
- Coverage measurement based on scenario-based simulation
- Online risk assessment