Researchers from the Cambridge Computer Vision for Digital Twins Research Group (CV4DT) led by Dr Olaf Wysocki, have had four papers accepted to NeurIPS 2026, spanning vision-language-action models, physics-informed digital twins, electromagnetic imaging, and physical reconstruction and simulation from video.
Together, the papers reflect a broad research interest in combining learning-based approaches with physical knowledge and structure to improve how complex systems can be modelled, reconstructed and understood.
Continually Evolving Skill Knowledge in Vision Language Action Model explores continual learning in vision-language-action models, introducing an approach that enables models to build and organise knowledge across tasks and skills while adapting to new manipulation tasks. arXiv
PhySPRING: Structure-Preserving Reduction of Physics-Informed Digital Twins via Graph Neural Networks investigates how graph neural networks can jointly perform system identification and topological reduction of spring-mass digital twins reconstructed from RGB-D videos. The aim is to reduce model complexity while retaining the physical behaviour needed for applications including robotic manipulation. alphaXiv
GauGal: Gaussian-Galerkin Electromagnetic Inverse Scattering Imaging explores how Gaussian primitives can serve both as a material representation and as the computational basis for solving electromagnetic wave scattering, enabling fast and physics-consistent reconstruction of an object’s internal permittivity from measured scattered fields.
PhysTwin++: Bending-Enhanced Differentiable Physical Reconstruction and Simulation from Videos focuses on the reconstruction and simulation of physical systems from video, extending work at the intersection of visual reconstruction and physics-based modelling.
The four papers demonstrate the breadth of research across CV4DT, bringing together artificial intelligence, computer vision, robotics and physics to explore challenges beyond traditional civil engineering applications.
Professor Brian Sheil said:
"This is only an outcome from doing the right things and an opportunity to, instead, celebrate the fantastic team, research culture and work ethic we have built over the last four years. What is especially exciting now is that we are tackling real civil engineering problems where we are seeing very strong interest from our industry partners, while also having the capability within the team to develop fundamental advances in robotics, physics-informed AI and subsurface reconstruction to address them. It is a tremendously exciting time for the Cambridge Laing O'Rourke Centre".
Congratulations to Dr Yixiong Jing, Dr Haibing Wu, Dr Guangming Wang, Dr Olaf Wysocki, Professor Brian Sheil and all the researchers and collaborators involved.