CREATE: Cross-Layer Resilience Characterization and Optimization for Efficient yet Reliable Embodied AI Systems

TL;DR

The paper tackles the energy-reliability bottleneck of embodied AI systems by introducing CREATE, a cross-layer resilience principle that jointly optimizes circuit (anomaly detection), model (weight-rotation), and application (entropy-driven voltage scaling) strategies. Through large-scale error-injection experiments, it demonstrates that planners are more vulnerable to timing errors than controllers, and that activation outliers and normalization skew in LLMs drive planner fragility. The authors implement a hardware-software co-design combining AD, WR, and VS on a systolic-array accelerator, achieving up to 40.6% computational energy savings vs nominal voltage and substantial chip-level and battery-life improvements across diverse tasks and platforms. This work enables practical, energy-efficient deployment of embodied AI by leveraging heterogeneous resilience to maintain reliability under aggressive voltage scaling. The approach is generalizable to various platforms and tasks, offering a concrete path toward durable edge AI systems.

Abstract

Embodied Artificial Intelligence (AI) has recently attracted significant attention as it bridges AI with the physical world. Modern embodied AI systems often combine a Large Language Model (LLM)-based planner for high-level task planning and a reinforcement learning (RL)-based controller for low-level action generation, enabling embodied agents to tackle complex tasks in real-world environments. However, deploying embodied agents remains challenging due to their high computation requirements, especially for battery-powered local devices. Although techniques like lowering operating voltage can improve energy efficiency, they can introduce bit errors and result in task failures. In this work, we propose CREATE, a general design principle that leverages heterogeneous resilience at different layers for synergistic energy-reliability co-optimization. For the first time, we conduct a comprehensive error injection study on modern embodied AI systems and observe an inherent but heterogeneous fault tolerance. Building upon these insights, we develop an anomaly detection and clearance mechanism at the circuit level to eliminate outlier errors. At the model level, we propose a weight-rotation-enhanced planning algorithm to improve the fault tolerance of the LLM-based planner. Furthermore, we introduce an application-level technique, autonomy-adaptive voltage scaling, to dynamically adjust the operating voltage of the controllers. The voltage scaling circuit is co-designed to enable online voltage adjustment. Extensive experiments demonstrate that without compromising task quality, CREATE achieves 40.6% computational energy savings on average over nominal-voltage baselines and 35.0% over prior-art techniques. This further leads to 29.5% to 37.3% chip-level energy savings and approximately a 15% to 30% improvement in battery life.

Figures (21)

• Figure 1: Embodied AI System Overview. (a) Embodied AI systems employ a hierarchical paradigm to tackle complex tasks: an LLM-based planner decomposes a task (often described in texts) into subtasks; then for each subtask, a low-level controller generates detailed actions step by step. These systems are typically battery-powered. (b) Lowering operating voltage introduces bit errors, causing (c) significant task performance degradation (e.g., lower task success rates and more task execution steps). (d) This in turn leads to higher energy consumption per task.
• Figure 2: CREATE Overview. CREATE is a general design principle to enable efficient yet reliable embodied AI systems. At the circuit level, anomaly detection and clearance mitigate large timing errors. At the model level, weight-rotation-enhanced planning redistributes LLM activations to improve robustness. At the application level, autonomy-adaptive voltage scaling dynamically adjusts voltage based on task demands. These techniques collectively boost both reliability and efficiency.
• Figure 3: JARVIS-1 Platform. JARVIS-1 consistently acquires high-level items in Minecraft playground. The planner generates a subtask sequence, while the controller integrates subtask prompts with visual observations to determine actions at each step. Both the planner and controller are primarily stacked Transformer blocks, which contain various network components, such as K and Down.
• Figure 4: Timing error model. (a) Bit-level timing error rate under different voltages. (b) Error pattern at 0.85V, overlapping with normal runtime activation distribution.
• Figure 5: Resilience Characterization. (a)-(b) Planner resilience characteristics. (c)-(d) Controller resilience characteristics. (e)-(f) Resilience comparison between K and O in the planner. (g)-(h) Resilience comparison between K and O in the controller. (i)-(j) Activation distributions of the pre-norm layer in the planner and controller. (k) Planner normalization outcomes exhibit significant skew under errors; (l) Controller normalization maintains moderate variance.