Robot skill acquisition, transfer and augmentation for indusrial automation

Project overview:
We are recruiting a self-funded PhD student to join an EPSRC project, STAMAN.
Though the studentship will not be provided, attending conferences, consumables, and industrial engagements will be funded and supported.

STAMAN's overview:
Many assembly and disassembly tasks in manufacturing have small clearances and limited accessibility, such as shaft-hole insertion/separation and bolt-nut assembly/disassembly. Using robots in these contact-rich tasks is more complex than those having no physical contacts (e.g. computer visual inspection) or simple contacts (e.g. cutting, welding, pick-and-place). The deployment of robots in contact-rich tasks has been limited to date. The contact-rich tasks that involve complex shapes, small clearances or deformable materials are particularly challenging to robotise due to the likely events of jamming and wedging.

Our previous research has investigated techniques that allow robots to learn contact-rich skills (e.g. complex motion plans and force control policies) using two main AI-based pathways: (1) self-learning from trial-and-error, and (2) learning from human demonstrations. The two participating universities, Birmingham and Sheffield, have research experiences in (1) and (2), respectively.

A key challenge observed in the current research is that in many cases a robot's contact-rich skill cannot be performed by other robots of different motion properties (e.g. accuracy, precision and stiffness), or be applied to a new task with variations (e.g. differences in object geometry, shape, and materials). This is because a robotic contact-rich skill, i.e. control policies and motion plans, is usually acquired for a specific task and cannot be adopted by new robots or in new tasks.

STAMAN's vision is to create AI-based mechanisms to allow robots to share and recreate obtained digital skills (e.g. motion and force/torque control strategies) to allow easy automation scale-up for contact-rich tasks. This includes considering two research questions:
1) For skill transfer - how can a contact-rich skill be quickly transferred to a different robot (e.g. transferring a bolt-nut separation skill from a high-precision robot to a low-precision robot)?
2) For skill augmentation - how can existing contact-rich skills be used to create new contact-rich skills (e.g. augmentation of rigid-material skills to deal with soft materials)?

The project will develop a portfolio of research into the science of digital skills for contact-rich tasks, focusing on common manufacturing tasks such as bolt-nut assembly/disassembly, peg-hole insertion/separation, and shaft-ring assembly/disassembly. The ability to transfer and augment digital skills for contact-rich tasks will allow automation systems to be implemented on a larger scale, with minimal manual setting and fine-tuning required. STAMAN aims to create transferrable and augmentable digital skills that will underpin the development of mass machine skills for future manufacturing, similar to how industrial robots have contributed to modern mass production.

The proposed research encourages more use of robots in assembly (e.g. automotive, aerospace, electronics, etc.) and disassembly (e.g. repairs, remanufacturing and recycling), and thus directly contributes to the UK's Made Smarter initiative and the circular economy goals.

2. Research Objectives
1. Develop and validate methods for robot system skill acquisition, transfer and augmentation
2. Practical development of new products and processes supported by industrial partners

3. Candidate Profile
A master’s degree (or equivalent) in engineering, computer science, applied mathematics, or a related discipline. Strong programming skills (e.g., Python). Knowledge of robotics, control, machine learning and embedded systems is desirable. Ability to work independently and engage with academic and industrial stakeholders.

References:

- Qu, M., Pham, D. T., Lan, F., Wu, Z., Zang, Y., Zhang, Y., & Wang, Y. (2025). Contact-Based Digital Twins Modelling for Reinforcement Learning of Robotic Disassembly Operations. IEEE Transactions on Industrial Cyber-Physical Systems.

- Zang, Y., Xu, X., Qu, M., Dixon, R., Ye, J., Hajiyavand, A. M., ... & Wang, Y. (2025). Robotic Disassembly Skill Acquisition Based on Reinforcement Learning With External Knowledge Injection. IEEE Transactions on Industrial Informatics.

- Deng, W., Liu, Q., Zhao, F., Pham, D. T., Hu, J., Wang, Y., & Zhou, Z. (2024). Learning by doing: A dual-loop implementation architecture of deep active learning and human-machine collaboration for smart robot vision. Robotics and Computer-Integrated Manufacturing, 86, 102673.

- Qu, M., Wang, Y., & Pham, D. T. (2023). Robotic disassembly task training and skill transfer using reinforcement learning. IEEE Transactions on Industrial Informatics, 19(11), 10934-10943.