Understanding Human Behaviour with Multimodal AI
Tanaya Guha
University of Glasgow | School of Computing ScienceFor machines to be considered truly intelligent, they must be capable of interacting with humans in a socially competent way. The key, therefore, is the machine's ability to recognise and interpret human’s behaviour (e.g., intent) and respond accordingly. Since human behaviour is inherently multimodal, our systems must be able to sense and fuse behavioural cues across multiple modalities. In this talk, I will discuss how to sense and model various verbal and non‑verbal behavioural cues to infer human's behavioural aspects, including internal states using deep multimodal models. Drawing on applications from domains such as mental health and wellbeing, as well as human–robot social interaction, I will outline the opportunities and challenges of multimodal AI in human behaviour understanding.
Reproducible Robotics for AI: Benchmark Failures and a Low-Cost Path Forward
Florent Audonnet and Gerardo Camarasa
University of Glasgow | School of Computing ScienceRobotics and robot learning are often showcased through highly polished demonstrations that conceal a persistent reproducibility and robustness gap between simulation and the real world. In this seminar, we will show how state‑of‑the‑art imitation and policy‑learning methods achieve high success rates in curated simulators yet collapse on slightly more complex, multi‑step, long‑horizon tasks, even before deployment to hardware. To address this problem, we introduce a low‑cost, 3D‑printed, track‑mounted robotic platform with a digital twin, designed for systematic evaluation across families of tasks such as pick‑and‑place, cube‑swapping, and constrained rearrangement to name but a few. Using this setup and recent transformer‑ and diffusion‑based policies, we quantify success, failure, and overfitting patterns under controlled task variations and horizon lengths. With this 3D‑printed setup, we will propose how standardised, affordable platforms can drive the development of robot learning approaches in real‑world scenarios.
