Peking University Researchers Achieve 30% Battery Life Extension for Embodied AI Systems with CREATE Framework

Breakthrough in Energy-Efficient [Embodied AI](/glossary/embodied-ai)

A team of researchers from Peking University, including Tong Xie from the School of Integrated Circuits and Yijiahao Qi, Jinqi Wen from the School of EECS, has achieved a significant breakthrough in energy-efficient embodied artificial intelligence. Their work addresses a critical challenge in deploying AI systems on portable, battery-powered devices: balancing energy efficiency with reliability.

The CREATE Framework

The team proposes CREATE, a novel cross-layer resilience framework that synergistically optimizes energy use and reliability. The framework achieves remarkable results:

• 40.6% computational energy savings compared to systems operating at nominal voltage
• 29.5% to 37.3% chip-level energy reduction
• 15% to 30% extension in battery life for [embodied AI](/glossary/embodied-ai) devices
• Maintains iso-task quality throughout the optimization

Three Key Innovations

CREATEs success lies in three core innovations that work together across system layers:

1. Anomaly Detection and Clearance at the Circuit Level: Suppresses large errors induced by timing violations, establishing a robust foundation for further optimizations
2. Weight-Rotation-Enhanced Planning at the Model Level: Redistributes LLM activations to improve robustness and maintain task quality while addressing persistent smaller errors
3. Autonomy-Adaptive Voltage Scaling at the Application Level: Dynamically adjusts the controllers operating voltage based on the current subtask execution status, maximizing efficiency

Discovering Heterogeneous Resilience

Through comprehensive error injection studies, the research team discovered that modern [embodied AI](/glossary/embodied-ai) systems exhibit inherent but heterogeneous fault tolerance across system layers. While both the LLM-based planner and [reinforcement learning](/glossary/reinforcement-learning) controller demonstrate good error robustness at low bit error rates (≤10⁻⁷), the controller displays significantly higher resilience at elevated BERs (10⁻⁷ to 10⁻³).

This key insight formed the foundation for CREATEs synergistic, cross-layer optimization strategy. The researchers also found that systematic activation outliers within the LLM planner, combined with normalization operations, contribute to its poor resilience at higher bit error rates.

Practical Implications

This research marks a significant step toward practical, robust, and efficient [embodied AI](/glossary/embodied-ai) agents capable of operating in real-world environments. The energy savings translate directly into longer operational times for battery-powered robots and autonomous systems used in industrial automation, search-and-rescue operations, and various service applications.

The team customized dedicated circuits for dynamic voltage scaling in systolic arrays and low-dropout regulators (LDOs) to holistically implement these optimizations, making CREATE not just a theoretical advancement but a deployable solution for next-generation embodied AI systems.