Elijah Almanzor is originally from the Philippines but moved to England when he was nine. He has always been interested in robotics and completed a Mechanical Engineering degree. Elijah enjoyed his degree, however it didn’t quite allow him to delve deeply in to robotics, which is one of the reasons he’s so excited to join the CDT. Elijah is looking forward to coming to Lincoln and enjoying the quintessential English town feeling. Elijah will be going on to study his PhD at the University of Cambridge and says it would, “be an honour to do my research at the University that Newton, Turing and Hawking went to”. He is currently very interested in designing more intelligent soft robots for soft/delicate harvest picking. In his spare time Elijah loves to play guitar, and has recently picked up skateboarding, so is looking forward to exploring the skateparks around Lincoln.
Deep Reinforcement Learning for Control of Robotic Manipulators for Grasping Strawberries in Simulation
The agricultural sector is under pressure from the exponential growth of human population, climate change and aging labour work-forces. Robotics and Autonomous Systems are therefore being researched as a possible solution to ensuring a sustainable global-food chain for the future. Intelligent robotic manipulators could be used to assist workers in operations such as harvesting, precision weeding and food handling in warehouses. This project will therefore look at the use of the Twin-Delayed Deep Deterministic Policy Gradient (TD3) Deep Reinforcement Learning (DRL) algorithm as an end-to-end control mechanism for the self-robotic learning of a policy capable of manipulating and handling randomised strawberry clusters, as harvesting strawberries require dexterous movements subject to unstructured configurations and chaotic dynamics.
Various research have already implemented DRL with robotic manipulators, however there are sparse works on the applications of DRL in the domain of agriculture. TD3 requires no predefined labelled training dataset as the low-level joint space control policy is learnt via agent interactions with the environment through hand-crafted reward functions. The project aims to learn the policy in simulation followed by transference to a real environment with additional research on reducing the reality-simulation gap.