Host: Jin-Peng Liu 刘锦鹏

Abstract:

Touch intelligence emerges from the interaction between a body, its environment, and an active sensing process. Unlike most visual sensing, tactile perception is intrinsically coupled to physical contact: the measured signal arises from deformation, stress transmission, friction, and relative motion at the sensor–object interface. Embodied touch must therefore be understood as a coupled problem spanning mechanics, optics, inverse modelling, probabilistic inference, and closed-loop control.

Using vision-based tactile sensing as a central example, this talk will examine the mathematical and physical foundations of embodied touch intelligence. I will describe how contact forces are transformed into observable images through elastic deformation, surface geometry, illumination, and camera projection. Inferring contact geometry and the underlying mechanical quantities, such as force distributions, frictional state, and material properties, from tactile images requires solving coupled inverse problems whose identifiability depends on the sensor model, calibration, loading conditions, and available temporal information.

The talk will introduce key models from continuum and contact mechanics, including stress–strain relationships, viscoelasticity, normal and tangential contact, friction-cone constraints, and the transition from sticking to partial and gross slip. It will also discuss how photometric stereo, marker displacement, optical flow, surface-normal integration, and inverse finite-element methods can be used to reconstruct contact geometry and force distributions. Particular attention will be given to observability, identifiability, regularisation, uncertainty, and the ambiguities that arise when different physical interactions produce similar tactile observations.

I will then examine sensor morphology as a form of mechanical computation. The geometry, stiffness, thickness, internal structure, and optical properties of a tactile sensor determine which spatial and temporal features are amplified, filtered, or lost before learning begins. From this perspective, tactile intelligence is not produced by an algorithm alone, but by the joint design of the physical body, sensing mechanism, inference model, and action policy.

Finally, I will discuss how physics-informed learning, multimodal representation learning, and active perception can connect these models to robotic manipulation. The talk will frame embodied tactile perception as a closed-loop process in which actions reduce uncertainty, test physical hypotheses, and maintain stable interaction. The central argument is that the next generation of touch-intelligent robots will require the co-design of mathematical models, material physics, sensor morphology, learning, and control.

Bio:

Dr Dandan Zhang is an Assistant Professor in Artificial Intelligence and Machine Learning at Imperial-X, Imperial College London. She leads the Multi-Scale Embodied Intelligence Lab. Her work investigates how physical interaction, sensor morphology, contact mechanics, optical perception, and machine learning can be integrated to enable robots to understand and manipulate the physical world across scales. In particular, her research on embodied touch intelligence connects the mathematics and physics of deformation, force transmission, inverse modelling, and multimodal perception with the design of advanced tactile sensors and adaptive robotic systems.

Dr Zhang was recognised as an AI Visionary in 2025 for her contributions as a female leader in AI for healthcare. Her distinctions include the IEEE ICRA Outstanding Interaction Paper Award, the UK Best PhD in Robotics Thesis Award, the Amazon PhD Prize for Outstanding Achievement in Robotics, and best-paper awards at IEEE MARSS and ICAC. Her work has also received multiple finalist nominations at IEEE IROS, MARSS, and Cyborg and Bionic Systems. Over the past five years, Dr Zhang and her research group have published more than 70 papers across leading journals and conferences, including Science Robotics, Communications Physics, Matter, Advanced Science, ACS Photonics, Advanced Optical Materials, Cyborg and Bionic Systems, IEEE Transactions on Robotics, IEEE Transactions on Automation Science and Engineering, IEEE Robotics and Automation Letters, ICRA, IROS, and MARSS.

https://www.intelligentrobotics-acrossscales.com