Research Themes
Ada-GO
We are creating a global optimization framework for co-evolving robot hardware (sensors, morphology, actuation, materials) and the reinforcement learning methods that control it. The goal is to make this co-evolution efficient through adaptive simulation, active learning, and hardware-in-the-loop optimization. By performing this search at a global scale, we aim to uncover novel, low-cost robot designs and powerful RL techniques that significantly outperform existing solutions. Ultimately, the aim is to drastically reduce the cost and time required to customize robots.
ARIA / Electrical Engineering News / Computer Science News / The Observer
Local Intelligence
Today’s autonomous systems are dependent on a stable connection to cloud servers, making them brittle, insecure, and inaccessible where connectivity is limited. Encode Fellow Martyna Stachaczyk is working with our team to compress large vision-language models for edge deployment and to design a fully on-device control architecture for real-time local intelligence. This research could free intelligent systems from the cloud, enabling safe, private, and adaptive autonomy for edge devices, even in resource-constrained or offline settings.
ARIA / Encode / LinkedIn Post
AIxSIM
With the emergence of various AI model scaling laws, a critical question arises: Can existing hardware sustain the continued growth of these models, or will a breakthrough architecture be required to deliver the same capabilities at radically lower costs? To explore such potentially transformative hardware designs, AIxSim—a project by teams from Imperial College London, the University of Edinburgh, and the University of Cambridge—aims to estimate their impact at both the system level (power and performance) and the model level (accuracy).