Towards adaptive human-AI and human-robot collaboration

Intelligent agents are increasingly being deployed in real-world scenarios where they interact and collaborate with humans, whether in the form of software agent chatbots or embodied intelligence in the form of cleaning robots, robot butlers and search and rescue robots. Hence, it is important for intelligent agents and robots to be able to robustly learn and adapt to preferences of any novel collaborative partners that they come across in real-world settings. However, in scenarios where agents or robots are to collaborate directly with humans, they require a significant amount of task relevant as well as human collaborative data. Collecting human collaborative data with intelligent agents and robots may not be practically possible. In addition, human preferences and skill sets may vary significantly, further increasing the challenge for generalizing over collaborative tasks.

This thesis discusses methodologies that allow intelligent agents and robots to robustly collaborate with humans with minimal access to human collaborative data. On the Human-AI Collaboration side, this thesis presents two works that focus on collaborative Reinforcement Learning (RL) agents. Firstly, we propose a Hierarchical RL agent that can dynamically adapt to novel collaborative partners by learning a two-level hierarchical policy. Secondly, we propose a methodology for generating collaborative agents with diverse preference in a computationally efficient manner using a learned World Model (WM) that simulates diverse collaborative trajectories. On Human-Robot Collaboration, we present a methodology that allows Visual Language Action (VLA) Models to learn collaborative tasks by autonomously generating collaborative robot trajectory data from natural language task prompts. Together, this thesis links adaptivity to novel partners as a crucial component in collaborative decision making with human partners across both AI agents and robots.

Speaker’s profile

Loo Yi is PhD candidate under the Information Systems Technology and Design under Singapore University of Technology and Design supervised by Prof Malika Meghjani. His research focuses on multi-agent and multi-robot systems, particularly how AI agents and robots can robustly collaborate and coordinate with human partners across a variety of tasks. He received his BEng also under ISTD pillar from SUTD in 2017.