Damage Handling Engineering For Soft Agricultural Robotics
Lead Research Organisation:
University of Cambridge
Department Name: Engineering
Abstract
The goal of the research project is to systematically investigate the potential engineering solutions in
hardware and software to a variety of challenges posed by hard-to-automate tasks in agricultural
robotics. The sample problem to be solved is how to effectively automate the selection and picking of
iceberg lettuces while minimising damage to the vegetable.
Approach
The approach will be to divide the project into four stages:
1. Study and understand the existing human-based solution
2. Break down the workflow into sub-problems and identify which can be solved with existing techniques
3. Use agile methodology to design, prototype and field test a complete autonomous agent, trying
different hardware and software solutions to the core problems
4. Summarise the research and document the final solution
A classic solution to the harvesting problem would include computer vision, some symbolic processing to take decisions and then inverse kinematics to control the gripper. The alternative approach proposed is to follow the design processes of embodied cognitive science which looks at a solution from the perspective of a complete autonomous agent. This gives great scope for the use of learning techniques inspired by
developmental processes in animals and humans. Rather than treat perception, decision and control as separate modules, the investigation will be how to let the agent learn the best solution to core problems by applying reinforcement and deep learning techniques over the combined modalities.
hardware and software to a variety of challenges posed by hard-to-automate tasks in agricultural
robotics. The sample problem to be solved is how to effectively automate the selection and picking of
iceberg lettuces while minimising damage to the vegetable.
Approach
The approach will be to divide the project into four stages:
1. Study and understand the existing human-based solution
2. Break down the workflow into sub-problems and identify which can be solved with existing techniques
3. Use agile methodology to design, prototype and field test a complete autonomous agent, trying
different hardware and software solutions to the core problems
4. Summarise the research and document the final solution
A classic solution to the harvesting problem would include computer vision, some symbolic processing to take decisions and then inverse kinematics to control the gripper. The alternative approach proposed is to follow the design processes of embodied cognitive science which looks at a solution from the perspective of a complete autonomous agent. This gives great scope for the use of learning techniques inspired by
developmental processes in animals and humans. Rather than treat perception, decision and control as separate modules, the investigation will be how to let the agent learn the best solution to core problems by applying reinforcement and deep learning techniques over the combined modalities.
Organisations
People |
ORCID iD |
| Fumiya Iida (Primary Supervisor) | |
| Simon Birrell (Student) |
| Description | The team has developed a lettuce-harvesting robot that can distinguish between mature and immature lettuces using a computer vision system and harvest the mature ones. This process has previously been completely manual. We have learned methods for calibration and handling of the vegetable. The current calibration and control methods are being revised to compensate for environmental noise. We are using a neural network and a revised visual servoing method for this. |
| Exploitation Route | The overall rate of damage (understood as failure to meet supermarket standards) is still too high and could be further improved. The device itself could be exploited commercially. |
| Sectors | Agriculture, Food and Drink |