This paper introduces a novel hardware-aware Neural Architecture Search (NAS) framework that integrates deployment-aligned low-precision training. This approach addresses the accuracy degradation caused by the mismatch between full-precision optimization and low-precision deployment on edge accelerators, particularly for spaceborne AI applications.

This paper introduces LSAI, a novel large and small AI model codesign framework, to enable agentic robots to perform accurate and real-time environment sensing and estimation with efficient path planning in complex scenarios like search and rescue. It aims to overcome limitations of traditional and singular large AI solutions in multi-robot cooperation by deeply integrating edge and terminal intelligence.

This paper introduces a practical framework for benchmarking and optimizing AI models on ARM Cortex processors in embedded systems. It focuses on balancing energy efficiency, accuracy, and resource utilization, demonstrating how optimal processor and model selections depend on an application's inference cycle time.