Title: Machine Learning techniques for domain adaptation of visuomotor representations for rapid deployment of pretrained robotic agents in new environments (short title: Bridging Simulation to Reality Gap in Robotics).
Funding: Full time Home/EU fees and a stipend of £15,285 p.a.
Application deadline: 20th January 2021.
Start date: April 2021
Duration: 3 years (full time)
Location: Colchester Campus
Deep convolutional networks, when trained on large-scale datasets, can learn representations which are generically useful across a variety of tasks and sensing domains.
However, due to a phenomenon known as dataset bias or domain shift, models trained along with these representations on one large dataset do not generalize well to novel datasets and tasks. The typical solution is to further fine-tune these networks on task-specific dataset.
Within industrial settings however, data gathering, and annotation is usually prohibitively expensive due to both technical (obtaining ground truth labels for negative examples would require operation under faulty condition) and non-technical reasons (e.g. data sensitivity).
One appealing alternative is producing synthetic data through simulations. For example, navigation and scene understanding models trained within robot simulators with 3D models, or defect detection algorithms trained on datasets produced by Finite Element Models would be able to be readily transferred on the corresponding real-world applications with principled fine-tuning.
This research will aim to provide a comprehensive set of techniques for supervised and unsupervised domain adaptation of pretrained robotic agents in new environments. The effectiveness of the developed methods will be demonstrated on well-benchmarked datasets and new tasks including visual object recognition and manipulation.
The overall work will produce open datasets and propose a new, open data standard incorporating methodologies to reduce domain shift in simulated datasets.
The research will have significant impact in accelerating the reduction of the exposure of personnel to hazardous environments, enhancing defect detection accuracy and reliability of safety-critical assets in inspection and maintenance industry.
This project is funded by Essex University, Lloyds Register Foundation and TWI.
The studentship includes:
Lead supervisorSchool of Computer Science and Electronic Engineering, University of Essex
Industrial supervisorDirector of Essex Artificial Intelligence Innovation Centre, Research and Enterprise Office
The successful candidate would be expected to speak fluent English and meet our English Language requirements.
Note: Overseas applicants should also submit IELTS results (minimum 6.5) if applicable
You can apply for this postgraduate research opportunity online.
Please upload your CV, personal statement, and transcripts of any undergraduate or masters programmes.
Instruction to applicants
When you apply online you will be prompted to fill out several boxes in the form:
If you have any informal queries about this opportunity please email supervisors; Dr Vishwanathan Mohan (lead supervisor) (firstname.lastname@example.org), Dr Panos Chatzakos (email@example.com) and Dr Delaram Jarchi (firstname.lastname@example.org).