Modelling robotic manipulation of deformable linear objects in constricting environments
In the UK alone, there are more than 30 automotive manufacturers building in excess of 70 models of vehicle. Together they produce Over 1.3 million cars and 78,000 commercial vehicles every year[1], [2]. Although already highly automated, the manufacturing of these vehicles still requires humans to install long, flexible, objects such as wiring, seals and hoses, and manufacturers have for a long time been looking for a way to automate this complex task.
The state of the art when it comes to installing such deformable linear objects or (DLOs) are human hands. To date, only humans have the dexterity and foresight to robustly handle DLOs in an industrial environment. With the advancement of sophisticated machine learning and autonomous control techniques in recent years, could DLO manipulation finally be automated?
The long-term vision of this research beyond that of this thesis is to have an autonomous robotic agent learn by itself how to manipulate DLOs around complex dynamic environments. The pipeline to make this vision a reality could consist of a Model Predictive Controller (MPC) used to guide a collaborative robot around an unstructured and dynamic environment towards a goal position while avoiding high force areas that can cause damage to DLOs. The MPC would use a learned recurrent model of the environment dynamics to plan safe trajectories to the goal. The robot would also use Reinforcement Learning (RL) to mostly follow the path set by the MPC but to maximize its RL reward, it may also explore areas around the path to and find more efficient routes to the goal. While the robot searches the environment task space, it generates new force data that is used to update the learned dynamics model, and the cycle continues until the most optimal path is found.
The scope of this thesis is a step towards this vision that focuses on the foundations needed by showing how a predictive dynamics model can be developed using advanced machine learning and data engineering to predict DLO pulling forces when such DLOs are pulled through and around constraining obstacles in dynamic environments.
The main outcome of this work was a predictive model capable of efficiently predicting manipulation forces over entire robot trajectories, even when the manipulated object was physically constrained by the workspace environment.
Results show that manipulation forces from bending and pulling DLOs can be efficiently modelled in 40 minutes using an LSTM with only a modest amount of training data.
Funding
Manufacturing Technology Centre
History
School
- Mechanical, Electrical and Manufacturing Engineering
Publisher
Loughborough UniversityRights holder
© Dominic McKeanPublication date
2023Notes
A Doctoral Thesis. Submitted in partial fulfilment of the requirements for the award of the degree of Doctor of Philosophy of Loughborough University.Language
- en
Supervisor(s)
Niels Lohse ; Peter KinnellQualification name
- PhD
Qualification level
- Doctoral
This submission includes a signed certificate in addition to the thesis file(s)
- I have submitted a signed certificate