CREATE: C ross-Layer Resilience Characterization and Optimization for Efficient yet R eliable E mbodied A I Sys te ms Conference: Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2; March 22–26, 2026; Pittsburgh, PA
Tong Xie, Yijiahao Qi, Jinqi Wen, Zishen Wan, Yanchi Dong, Zihao Wang, Shaofei Cai, Yitao Liang, Tianyu Jia, Yuan Wang, Runsheng Wang, Meng Li · 2026
Read CREATE: C ross-Layer Resilience Characterization and Optimization for Efficient yet R eliable E mbodied A I Sys te ms Conference: Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2; March 22–26, 2026; Pittsburgh, PA by Tong Xie, Yijiahao Qi, Jinqi Wen, Zishen Wan, Yanchi Dong, Zihao Wang, Shaofei Cai, Yitao Liang, Tianyu Jia, Yuan Wang, Runsheng Wang, Meng Li (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14140.txt.
Introduction
Tong Xie and a large team harden embodied AI systems against faults across design layers in 2026. Embodied AI bridges AI with the physical world. Modern systems often combine a language model planner for high-level tasks. A reinforcement learning controller handles low-level action generation in the same stack.
Their CREATE framework characterizes resilience across layers and optimizes it. The aim is efficient yet reliable embodied AI systems. Faults can enter through sensors, planners, controllers, or hardware. Cross-layer views catch weak points that single-layer checks miss. Each layer gets its own checks, from the lens to the plan. Weak spots that hide in one layer show up when the layers are read as one.
Catalog record
| CREATOR | Tong Xie, Yijiahao Qi, Jinqi Wen, Zishen Wan, Yanchi Dong, Zihao Wang, Shaofei Cai, Yitao Liang, Tianyu Jia, Yuan Wang, Runsheng Wang, Meng Li |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-14140
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CREATE: C ross-Layer Resilience Characterization and Optimization for Efficient yet R eliable E mbodied A I Sys te ms Conference: Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2; March 22–26, 2026; Pittsburgh, PA, USA Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’26), March 22–26, 2026, Pittsburgh, PA, USA DOI: 10.1145/3779212.3790147 ISBN: 979-8-4007-2359-9/2026/03 CCS: Hardware Robustness CCS: Computer systems organization Architectures CCS: Computer systems organization Embedded and cyber-physical systems Tong Xie, Yijiahao Qi, Jinqi Wen, Zishen Wan, Yanchi Dong, Zihao Wang, Shaofei Cai, Yitao Liang, Tianyu Jia, Yuan Wang, Runsheng Wang, Meng Li 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. † † cc-license: by 1. Introduction Embodied artificial intelligence (AI) is a class of intelligent systems capable of understanding, reasoning, and interacting with the physical world ( 59 ; 21 ; 98 ) . Distinguished by their proficiency in executing complex long-horizon, multi-objective tasks, these systems have demonstrated broad application potential across various domains, including industrial manufacturing ( 11 ) , logistics and transportation ( 24 ) , and search-and-rescue operations ( 105 ) . As shown in Fig. 1 (a), to achieve such capabilities, embodied agents typically employ large language models (LLMs) or multi-modal language models (MLMs) as planners to decompose the entire tasks into specific subtasks , alongside reinforcement learning (RL)-based controllers to determine low-level actions for each subtask, which typically requires thousands of steps ( 3 ; 20 ; 79 ; 92 ; 93 ; 9 ; 11 ) . Task performance is usually measured by the success rate and average completion steps ( 93 ; 39 ; 79 ; 11 ; 92 ; 9 ) . Unlike traditional robotic systems driven by predefined rules ( 80 ; 25 ) , embodied agents leverage large-scale LLMs, which would often consume trillions of operations per inference. Although cloud-based inference via APIs enables access to giant LLMs, e.g., GPT-4 ( 1 ) , it introduces high communication latency and reliance on Wi-Fi, limiting desirable real-time interaction capabilities for many applications. Therefore, recent research has begun to explore the end-to-end deployment of embodied agents on edge hardware, such as GPUs ( 94 ; 12 ; 20 ) and customized accelerators ( 91 ; 3 ; 39 ; 11 ; 83 ) , such as TPU-like systolic arrays ( 43 ; 64 ) . 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. However, such edge deployment of embodied AI systems is often strictly energy-constrained due to reliance on battery power ( 59 ; 65 ; 91 ; 83 ; 23 ) . With intensive computational workloads (e.g., long-context LLM inference and high-frame-rate image processing), these computation platforms have been extensively reported to consume substantial power ( 89 ; 48 ; 11 ; 20 ; 53 ; 39 ; 84 ; 80 ; 101 ; 40 ) . For example, ( 40 ) reported that the computing platform accounted for over 50% of total power on quadruped robots ( 4 ) , and ( 39 ) demonstrated that LLM inference dominated the total energy consumption in a representative robotic setup using an NVIDIA GPU and Franka robot arm ( 26 ; 53 ; 48 ) . Therefore, it is imperative to improve the power efficiency of embodied AI systems. Given the quadratic relation between power consumption and operating voltage, directly reducing the voltage can improve processing efficiency. This method, however, can induce timing violations ( 19 ; 67 ; 42 ; 38 ) and increase computation bit error rates (BERs), as shown in Fig. 1 (b), leading to severe degradation of task performance, such as lower task success rates and more execution steps (Fig. 1 (c)). These can ultimately increase the energy consumption to complete a single task (Fig. 1 (d)). Various techniques have been proposed to balance reliability with efficiency at the circuit, algorithm, and system levels ( 54 ; 46 ; 81 ; 103 ; 102 ; 69 ; 22 ; 28 ; 95 ; 37 ; 8 ; 73 ; 99 ; 76 ; 50 ; 32 ; 88 ; 41 ; 90 ; 77 ) , but all remain prohibitively costly or impractical for embodied AI systems. Unlike traditional robotics, modern embodied agents are predominantly based on AI and deep neural networks (DNNs), which exhibit inherent error resilience ( 52 ; 72 ; 63 ; 87 ; 2 ; 88 ; 97 ) . For example, as in Fig. 1 (c), task success rates ( stone in ( 93 ) ) do not immediately degrade under low BERs, suggesting opportunities to save energy by strategically reducing the operating voltage without compromising the task performance. However, the error resilience of embodied AI systems is not well characterized yet, making it hard to determine a proper voltage scaling policy. While prior studies have examined the resilience of individual