Learning to collaborate: division of labour, human modelling and embodiment in intelligent systems

Read full text Summarise with VIRAT
Authors: Ruaridh Mon-Williams
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This thesis investigates topics that constitute key scientific milestones towards building robots able to effectively collaborate with humans in dynamic environments (e.g., healthcare, assisted living, and cooking). It focuses on increasing our knowledge about three central challenges to achieving this goal: (i) understanding the context-dependent division of labour in humans, (ii) modelling and predicting human behaviour, and (iii) improving the ability of robots to interact with the physical world in uncertain settings. Five studies are presented. The first two studies use the Overcooked-AI platform to examine how task concurrency influences whether agents specialise or act independently. A closed-form bound based on Amdahl’s Law predicts when specialisation improves performance. Experiments were conducted with both AI agents and human teams across randomised layouts. The results show that there are underlying principles that govern how humans and AI agents collaborate, and demonstrate that human behaviour closely aligns with the model’s predictions. Studies 3 and 4 explore how AI agents can model and adapt to humans. Study 3 introduces BeTrans, a transformer-based framework enabling rapid adaptation to non-stationary human behaviours, outperforming standard techniques. Study 4 shows that partner modelling can emerge spontaneously in model-free agents under appropriate social pressures. Study 5 addresses physical interactions in robotic systems. A GPT-4-based Embodied Language Model system (ELLMER) is developed to perform long-horizon tasks under uncertainty by integrating language-based planning with force and visual feedback. This framework enables a robot to execute complex tasks such as making coffee, representing progress toward scalable, embodied intelligence. These contributions support the development of collaborative intelligence. Analysing tasks and environments through parallel computing reveals principles that shape labour division in human and multi-agent systems. Behavioural prediction (with insight into the mechanisms enabling partner modelling) supports adaptation to diverse agents, while embodiment enables skilled interaction with the physical world. Integrating cognitive science, robotics, and multi-agent learning, this thesis offers insights into designing artificial agents that collaborate adaptively, flexibly, and intelligently in complex environments.

Fields of Study & Subjects

Robot Manipulation and Learning Social Robot Interaction and HRI Reinforcement Learning in Robotics

Publication details

Journal / source
ERA
Publication date
25 March 2026
Language
EN
Document type
Thesis