Robot Vision, Deep Learning, and Explainable AI.
- Rogers, H. and Fox, C. (2020) ‘An Open Source Seeding Agri-Robot‘, UKRAS20 Conference: Robots into the real world Proceedings.
- Rogers, H., Dawson, B., Clawson, G., and Fox., C. (2021) ‘Extending an Open Source Hardware Agri-Robot with Simulation and Plant Re-identification‘, Oxford Autonomous Intelligent Machines and Systems Conference 2021.
- UKRAS Conference (2020): An Open Source Seeding Agri-Robot.
- ICRA Task-Informed Grasping Workshop – III (2021): Robotic Manipulators in Agriculture: A Brief Review, watch here.
- Joint Robotics CDT Conference in Robotics & Autonomous Systems 2021: Extending an Open Source Hardware agri-robot with simulation and plant re-identification.
- AgriFoRwArdS CDT Annual Conference (2022): Explainable Droplet Recognition System for Precision Sprayer Applications.
Other Activities and Outputs
- Member of the AgirFoRwArdS CDT Drink Outside the Box Organisation Committee.
- Took part in the AgriFoRwArdS Summer School 2021 resulting in a co-authored presentation at the AgriFoRwArdS CDT Annual Conference 2021: RAS Counter: Non-invasive yield prediction for vineyards (in collaboration with Karoline Heiwolt, Ni Wang, Obinamuni De Silva, Huijiang Wang).
My name is Harry, I joined the CDT in September 2020. Before this I was at the University of Lincoln completing a BSc in Computer Science looking to work in robotic software development. I am interested in robot vision, deep learning and robot navigation. During my undergraduate I published a paper about my dissertation project in which I built and programmed an agricultural robot, which dispensed seeds and was tracked via GPS.
During the MSc Robotics and Autonomous systems, I have worked part time on the BACCHUS project. This has been fun to be work on an actual project that will be deployed to have a system in a vineyard. I have also worked on a paper for ICRA Task-Informed Grasping Workshop with members from the CDT as well as the MSc. Finally, I have also helped setup the Drink Outside the Box as a member of the committee
An Empirical Comparison of Optimisation Methods for Embedded DNNs
Automated precision agriculture is imperative and needs help to be optimised. To enable this Deep Neural Networks (DNNs) need to be deployed to ensure precision is kept high. Deploying DNNs like a Faster R-CNN to complete object detection can outperform YOLO DNNs, however these are much more difficult to deploy due to the size of the DNN. This thesis completes an empirical comparison of optimisations and considers deployment on multiple different embedded devices. This thesis completes multiple types of testing on multiple types of embedded devices with differing backbones for a Faster R-CNN. This thesis discovers that when DNNs are deployed not all optimisations can optimise DNNs for embedded deployment. This thesis also finds that each embedded device tested had optimal results with different combinations of optimisations. This thesis also contributes the usage of multiple quantisation optimisations for a Faster R-CNN. When using the multiple quantisation methods there can be seen a DNN size reduction of 52.4% and 67% with accuracy increases of 0.2% and 0.3% for a MobileNetV3-Large backbone and ResNet18 backbone, respectively. The test data is object detection based and can be used for grape harvesting within agricultural robotics.
Closing The Loop on Precision Spraying.
Despite the very long history of pesticides and herbicides it was not really until the world war II end that there was widespread use of synthesised chemicals in agriculture. Since then, there have been considerable developments often due to unintended or modelled consequences which usually amount to toxicity to animals or organisms essential to the ecosystem. The modern farmer is expected to be on top of considerable advice and guidance. Despite these advances the modern approach is known as “spray and pray” to imply that however carefully the diluted the chemicals are, there are always unintended consequences. A much-touted alternative is precision spraying in which exactly the right toxin is delivered to exactly the right place at exactly the right time. The technology already exists for the delivery of precise quantities of fluids through jets or aerosols. The question is, how feasible is it to monitor and control such systems in the field? That is what this project is about.
This project aims to incorporate improvements in Syngenta’s existing platforms for autonomous targeting of crops, weeds, pests, accurate dispensing of fluid. We will formulate quantifiable methodologies for post spraying effect monitoring to help Syngenta fulfil the regulatory guidelines for EU green deal. We will get access to the in-field/test equipment and relevant data (video/field maps) to build realistic and practical solutions in terms of precision spraying.