In the realm of artificial intelligence and energy efficiency, a team of researchers from the University of California, Berkeley, and the University of Texas at Austin have made significant strides. Led by Tong Xie and including Yijiahao Qi, Jinqi Wen, and others, the team has developed a novel approach to optimize the energy efficiency and reliability of embodied AI systems. Their work, titled “CREATE: Cross-Layer Resilience Characterization and Optimization for Efficient yet Reliable Embodied AI Systems,” was recently published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies.
Embodied AI systems, which integrate AI with the physical world, often combine large language models (LLMs) for high-level task planning and reinforcement learning (RL) controllers for low-level action generation. These systems are crucial for tasks in real-world environments but face challenges due to high computational demands, particularly for battery-powered devices. While techniques like lowering operating voltage can improve energy efficiency, they can also introduce bit errors and lead to task failures.
The researchers propose CREATE, a design principle that leverages heterogeneous resilience at different layers for synergistic energy-reliability co-optimization. They conducted a comprehensive error injection study on modern embodied AI systems and observed an inherent but heterogeneous fault tolerance. Based on these insights, they developed an anomaly detection and clearance mechanism at the circuit level to eliminate outlier errors.
At the model level, the team proposed a weight-rotation-enhanced planning algorithm to improve the fault tolerance of the LLM-based planner. Additionally, they introduced an application-level technique called autonomy-adaptive voltage scaling, which dynamically adjusts the operating voltage of the controllers. The voltage scaling circuit is co-designed to enable online voltage adjustment.
Extensive experiments demonstrated that CREATE achieves significant computational energy savings without compromising task quality. On average, it achieved 40.6% computational energy savings over nominal-voltage baselines and 35.0% over prior-art techniques. This led to 29.5% to 37.3% chip-level energy savings and approximately a 15% to 30% improvement in battery life.
For the energy sector, this research offers promising applications. Embodied AI systems are increasingly used in energy management, smart grids, and renewable energy integration. The CREATE framework can enhance the energy efficiency and reliability of these systems, leading to more sustainable and cost-effective energy solutions. By optimizing the energy consumption of AI-driven energy management systems, the framework can contribute to reducing the overall energy footprint of the sector.
The practical implications of this research are vast. Energy companies can deploy more efficient and reliable AI systems for monitoring and managing energy infrastructure. Smart grids can benefit from AI controllers that operate more efficiently, reducing energy losses and improving grid stability. Additionally, the framework can be applied to renewable energy systems, such as wind and solar farms, to optimize their performance and reliability.
In conclusion, the CREATE framework developed by the team of researchers offers a significant advancement in the field of embodied AI systems. Its potential applications in the energy sector can lead to more efficient and reliable energy management, contributing to a more sustainable energy future. The research was published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, providing a robust foundation for further exploration and development in this area.
This article is based on research available at arXiv.