Allowing autonomous robots to continually learn, generalize, and improve from their experiences

About the Project

To perform autonomously tasks such as object rearrangement, assembly, manipulation, and navigation, robots must be able to plan their actions over long horizons. Such planning is usually computationally challenging to perform in real time, especially considering complex robots and task specifications, or large and uncertain planning domains, with many irrelevant objects and distractions. One intuitive approach supporting autonomous robots in this challenge is by allowing them to learn to continually improve their planning capabilities over time, based on their experience. This general approach should enable us to build long-lived, multi-purpose robots with human-like versatility and common sense, rather than highly specialized machines.

Unfortunately, despite recent advancements in Machine Learning and "Learning from Demonstrations," existing learning approaches are not suitable for this objective, as these require numerous annotated demonstrations, rendering them unsuitable for online, autonomous learning.

To this end, our recent work introduced a novel algorithmic framework for automatic learning of "planning strategies" by abstracting successful planning experiences. This framework allows a robot to automatically and continually make generalizable conclusions from individual experiences, which can later be adapted for and reused in new contexts, to accelerate the solution of new planning problems - just like humans do, but without human intervention!

Initial results demonstrated the potential of this approach to significantly impact the field of AI-enabled robotics. To achieve that, this project seeks to extend this initial effort in various directions, including: application and adaptation to new platforms, planning domains, and task-types; application to multi-robot and human-robot collaborative systems; integration with (statistical) Machine Learning and Computer Vision techniques, Control, knowledge graphs and other components in the autonomy stack; improving utility and computational tractability through algorithmic development; and improving trustworthiness through formal analysis.

The work on this project is diverse and contains theoretical, computational. and experiential aspects. Students are expected to conduct research, publish papers, develop and release open-source code, and work with physical robots. You will have access to state-of-the-art hardware and resources, and excellent mentorship. Potentially, successful students will have access to collaboration and internship opportunities with industry leaders, such as NASA Robotics, Amazon Robotics and Bosche.

While prior research experience in robotics is recommended, it is not mandatory. Excellent candidates with background in robotics, AI, computer science, algorithms, applied mathematics, engineering, or other relevant backgrounds are welcome to apply.

References

www.khen.io

[1] Accelerating Long-Horizon Planning with Affordance-Directed Dynamic Grounding of Abstract Strategies, by Khen Elimelech, Zachary Kingston, Wil Thomason, Moshe Y. Vardi, and Lydia E. Kavraki, in IEEE International Conference on Robotics and Automation (ICRA), May 2024.

[2] Extracting generalizable skills from a single plan execution using abstraction-critical state detection, by Khen Elimelech, Lydia E. Kavraki, and Moshe Y. Vardi, in IEEE International Conference on Robotics and Automation (ICRA), May 2023.

[3] Principles of Robot Motion: Theory, Algorithms, and Implementations, by Howie Choset, Kevin M. Lynch, Seth Hutchinson, George Kantor, Wolfram Burgard, Lydia E. Kavraki and Sebastian Thrun, The MIT Press, 2005

Eligibility Criteria

While prior research experience in robotics is recommended, it is not mandatory. Excellent candidates with background in robotics, AI, computer science, algorithms, applied mathematics, engineering, or other relevant backgrounds are welcome to apply.

Application Process

Applicants are encouraged to contact Dr Khen Elimelech at (khen.elimelech@kcl.ac.uk) before formally submitting their application, including a summary of their academic background, research interests, and any relevant previous experience.

Once you have spoken with Dr Elimelechand you are both happy to proceed, please follow the formal application process below.

Formal Application

To be considered for the position, candidates must apply through the King’s Apply online application system. Full details are available on the Research degrees | Department of Informatics | King’s College London

When applying, please select:

• Programme: Computer Science Research MPhil/PhD (Full-time)
• Supervisor: Dr Khen Elimelech
• Project title: Include this in your application and all correspondence

In the Funding section of the form:

1. Select Option 5 – “I am applying for a funding award or scholarship administered by King’s College London.”
2. Enter the following code exactly as written in the *‘Award Scheme Code or Name’* box: ‘462’

Funding Notes

A fully funded PhD studentship for 3.5 years to start in October 2026.

Stipend: roughly £22,780.00 per annum, subject to increase

Bench Fees: £1,000.00 per annum

Tuition fees: covered in full for home/overseas students