Survey and Roadmap for Efficiency-Accuracy Co-optimization Towards Embodied AI Chip Design
Authors: Zhao, J., Yu, H., Li, Y.
Journal: Proceedings of Isoirs 2026 Moving Towards Embodied Intelligence in the AI Age 2026 6th International Symposium on Intelligent Robotics and Systems
Publication Date: 01/01/2026
DOI: 10.1109/ISOIRS70157.2026.11545322
Abstract:The embodied intelligence chip is the “brain” of the embodied agent, which enables the agent to autonomously make decisions and perform actions in complex and dynamic real physical environments. Conventional chip design optimization methodologies are predominantly based on static solutions and manual strategies. These methods face the “curse of dimensionality” and time-varying challenges such as dynamic loads and thermodynamic constraints. This article reviews related research in three aspects: dynamic-aware chip architectures, intelligent optimization strategy design, and LLM-driven automated design. For prospects, the “AI co-design paradigm” is constructed, in which AI is not only specialized in design, but also in requirement analysis, multiscale modeling, knowledge transfer, strategy enhancement, and verification, to realize autonomous design optimization. Jointly, outlook is drawn with respect to constructing MCU+NPU heterogeneous computing architectures, SSM world model collaboration mechanisms, and integrated sensing-computing-control systems.
Source: Scopus