Research

Our research deals with the design of embedded computing and communication systems and their underlying hardware and software technologies.

Energy-efficient and adaptive AI systems: Power and efficiency are critical design concerns across a wide range of platforms, from tiny IoT devices to large-scale cyber-physical systems. Our work explores hardware–software co-design of system architectures, approximation techniques, and adaptive execution strategies that significantly reduce energy consumption while maintaining high accuracy. Recent projects include ECO, a context-aware multimodal AI framework for NPUs, and XAMBA, which enables efficient state-space models on resource-constrained accelerators.

Battery-free and energy-harvesting embedded platforms: Energy harvesting opens the door to sustainable and perpetual system operation but requires carefully balancing energy availability and workload execution. We investigate novel computing paradigms such as processing-in-sensor and processing-in-memory that eliminate costly data movement and enable AI in battery-free devices. Our PRISM framework demonstrates how these architectures support energy-efficient AIoT in intermittently powered systems.

Efficient machine learning on NPUs: As emerging models such as graph neural networks and state-space models gain traction, their deployment on resource-constrained neural processing units requires new techniques. Our work, including GraNNite and XAMBA, develops high-performance execution frameworks for such models under strict compute and memory budgets.

Collaborative and distributed inference: Modern embedded systems often operate as part of a network of devices. We design multisystem approximation techniques that enable devices to jointly execute AI workloads, trading off accuracy, robustness, and energy efficiency in a coordinated manner.

Secure and reliable embedded systems: Reliability and security are vital in autonomous and connected platforms. Our SecuPilot project integrates a dedicated security coprocessor for UAVs, ensuring trustworthy and efficient operation. We continue to explore secure-by-design architectures and system-level reliability for next-generation embedded intelligence.