Research

At the core of our research lies the ability of the robot system to perceive what humans are doing and infer what their internal cognitive states (including beliefs and intentions) are. We use multimodal signals (RGB-D, event-based (DVS), thermal signals, haptic and audio) to perform this inference. We research most pipeline steps of human action perception (eye tracking, pose estimation and tracking, human motion analysis, and action segmentation). We collect and publish our datasets for the community’s benefit (see section Software on this website).

We are interested in learning and maintaining multiscale human models to personalise the contextual and temporal appropriateness of the assistance our robots provide – “how should we help this person, and when”. We typically represent humans at multiple levels of abstraction: from how they move, i.e., their spatiotemporal trajectory representations using statistical and neural network representations, to how they solve tasks, i.e., their action sequences using context-free grammatical representations, and how they use assistive equipment. Thus, we term our models “multiscale” models.

We are researching the learning processes that allow humans and robots to acquire and represent sensorimotor skills. Our research spans several representational paradigms: from embodiment-oriented statistical and neural representations, to more cognition-oriented ontological and knowledge-graph-based representations. We use active learning (motor and goal babbling and exploration) and social learning (e.g. human observation and imitation) to expand the range of skills our robots have.

How can robots plan and execute actions in an optimal, safe and trustworthy manner? We research robot motion generation algorithms in individual and collaborative tasks, paying particular attention to how control can be shared between human and robot collaborators. Apart from fundamental issues in human-robot shared/collaborative control, we are also interested in the interaction between multiple humans and robots, for example in triadic interactions (for example, an assistive robot, an assisted person and their carer)

Virtual and Augmented Reality Interfaces have great potential for enhancing the interaction between humans and complex robot systems. Our research investigates how visualising and interacting with mixed reality information (for example dynamic signifiers, or dynamically- determined affordances) can facilitate human collaboration through enhanced explainability, and fluidity and efficiency of control.

As robots become more integrated into our lives, it's crucial they understand and respect privacy and trust. We focus on three key areas: 1) enabling robots to gauge human trust and adapt their behaviour accordingly, 2) designing robots with clear and understandable decision-making processes, and 3) ensuring they learn personalised behaviours without compromising user privacy.

Caregiving robotics, as a subset of assistive robotics, are developed to provide support for care-related tasks for elderly individuals and people with mobility impairments, aiming to improve the quality of their lives and independence while also reducing the workload on human caregivers.