🤖 AI

XiYOLO: Energy-Aware Object Detection via Iterative Architecture Search and Scaling

XiYOLO introduces an energy-aware object detection framework that combines iterative neural architecture search with a two-stage energy estimator and compound scaling to generate a family of models achieving superior accuracy-energy tradeoffs on heterogeneous edge devices compared to YOLO baselines.

Original authors: Tony Tran, Richie R. Suganda, Bin Hu

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tony Tran, Richie R. Suganda, Bin Hu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to fit a high-performance camera system onto a small, battery-powered drone. You want the drone to see everything clearly (high accuracy), but you also need it to fly for a long time without the battery dying (low energy).

The problem is that most "smart" camera systems are like heavy, gas-guzzling trucks: they see well, but they drain the battery fast. Smaller systems are like bicycles: they save energy, but they might miss important details or struggle in the dark.

This paper introduces XiYOLO, a new way to design these camera systems so they are the "perfect hybrid vehicle" for drones and robots. Here is how they did it, broken down into simple concepts:

1. The Problem: The "One-Size-Fits-None" Dilemma

Usually, engineers have to pick one specific camera model for a specific drone. If they want a bigger drone, they have to design a whole new camera from scratch. Also, guessing how much battery a camera will use is like trying to guess how much gas a car will use just by looking at its engine size; it's actually about how the engine moves, the weight of the car, and the road conditions. Real-world energy is hard to predict without actually testing it, which takes forever and uses up the battery you are trying to save.

2. The Solution: The "Master Blueprint" (XiYOLO)

Instead of designing a new camera for every single drone, the authors created a Master Blueprint (called a "base architecture"). Think of this like a high-quality, adjustable tent.

The Search: They didn't just guess the best design. They used a smart computer program (called "Neural Architecture Search") to test thousands of tiny variations of the tent's poles, fabric, and zippers.
The XiResOFA Block: This is the special "building block" they invented. Imagine a Lego brick that can change its shape. It can be thick or thin, have a wide or narrow view, and use a simple or complex lens. The computer tested millions of these Lego combinations to find the one that sees the best while using the least energy.

3. The "Two-Stage Energy Estimator": The Smart Guessing Machine

Since they couldn't test every single design on every real drone (it would take too long and drain too many batteries), they built a two-stage energy predictor.

Stage 1 (The General Rule): A smart model that knows, "Generally, bigger lenses use more energy."
Stage 2 (The Specific Fix): A tiny, quick adjustment that says, "But on this specific drone, the battery behaves a little differently, so let's tweak the guess."
The Result: This allowed them to find the perfect design using only a tiny handful of real-world tests (like tasting a soup with just 2–20 spoonfuls instead of the whole pot).

4. The "Compound Scaling": Stretching the Blueprint

Once they found the perfect "Master Blueprint," they didn't stop there. They used a technique called Compound Scaling to create a whole family of cameras from that one blueprint:

Nano/Small: Stretch the blueprint slightly to make a tiny, super-efficient camera for small drones.
Medium: Stretch it a bit more for bigger drones that need to see further.
The Magic: Because they started with one perfect blueprint, all these different sizes work well together. You don't need to redesign the engine for every car size; you just scale the same efficient design.

5. The Results: Faster, Smarter, and Longer-Lasting

When they tested these new cameras on real drones (specifically the ModalAI Sentinel drone with its GPU and NPU chips):

Better Vision: On the PascalVOC test, the "Medium" XiYOLO camera saw objects slightly better than the current best (YOLOv12).
Huge Energy Savings:
- On the GPU (the drone's main brain), it used 20.6% less energy than the competition.
- On the NPU (a specialized chip for AI), it used 35.9% less energy.
Real-World Impact: In a long flight test, the XiYOLO cameras saved enough battery to fly significantly longer than the old cameras. For example, on the NPU, the new camera saved about 165 Joules of energy compared to the best old model over the course of the test.

Summary

The authors didn't just build a better camera; they built a system to design better cameras that adapts to different battery limits. They found a "Goldilocks" design that isn't too big or too small, and they proved that by using a smart guessing machine, you can find these perfect designs without draining your batteries in the lab. The result is a family of cameras that let drones see clearly while flying longer.

Technical Summary: XiYOLO: Energy-Aware Object Detection via Iterative Architecture Search and Scaling

Problem Statement

Object detection on heterogeneous edge devices (e.g., UAVs, UGVs, mobile robots) faces a critical conflict between the need for reliable perception and strict constraints on energy, latency, and memory. While larger detectors often improve accuracy, they incur prohibitive energy costs that limit system endurance. Conversely, smaller models are efficient but may lack robustness.

Existing energy-aware Neural Architecture Search (NAS) methods often target limited deployment settings or single final models rather than scalable detector families. A primary bottleneck is the difficulty of optimizing for "real" energy: standard proxies like FLOPs, MACs, or parameter counts fail to predict measured energy accurately because energy consumption is highly device-dependent, influenced by memory movement, operator composition, and platform-specific execution behaviors. Furthermore, obtaining exhaustive hardware energy measurements across a large search space is costly and often impractical.

Methodology

The authors propose XiYOLO, an energy-adaptive framework designed to identify a single energy-efficient base architecture and scale it into a family of detectors across different deployment budgets. The framework consists of four core components:

1. Energy-Aware Search Space (XiResOFA)

The framework introduces a novel search block, XiResOFA, which augments standard residual bottleneck designs with elastic architectural choices compatible with Once-for-All (OFA) search. Each block is parameterized by a tuple $(r, k, t)$ :

Compression Ratio ( $r$ ): Selectable from $\{1, 0.5, 0.25\}$ to control channel width.
Kernel Size ( $k$ ): Selectable from $\{1, 3, 5\}$ to adjust receptive fields.
Attention Type ( $t$ ): Selectable between "lite" and "full" attention mechanisms.

These choices are applied consistently across the backbone, Feature Pyramid Network (FPN), and Path Aggregation Network (PAN), creating a coherent search space where architectural decisions have a direct, interpretable relationship to both detection quality and hardware energy.

2. Two-Stage Energy Estimation

To address the scarcity of hardware energy data, the authors develop a two-stage energy approximation model:

Stage 1 (Generic Prior): A pretrained predictor $f_{base}(a)$ estimates energy based solely on architecture, capturing transferable structure-level energy patterns.
Stage 2 (Device-Specific Residual): A lightweight residual model $r(a, h)$ adapts the prior to specific hardware targets ( $h$ ), accounting for backend implementation and memory hierarchy.
The final estimate is $\hat{E}(a, h) = f_{base}(a) + r(a, h)$ . This approach allows the system to calibrate for a new device using only 2–20 measured samples (few-shot adaptation) rather than exhaustive profiling.

3. Iterative Architecture Search

Instead of a monolithic joint search over the backbone, FPN, and PAN, the framework employs an iterative search strategy. It progressively refines each component while conditioning on the current choices of the others:

Optimize the backbone given current FPN/PAN.
Optimize the FPN given the new backbone and current PAN.
Optimize the PAN given the new backbone and FPN.
This stage-wise refinement reduces combinatorial complexity while allowing components to co-adapt to the energy budget.

4. Compound Scaling

Once a single optimized base architecture ( $a^*$ ) is identified, the framework applies compound scaling to generate a family of deployable variants (nano, small, medium). This separates architecture discovery from deployment adaptation, enabling the generation of models for different operating regimes without requiring separate searches for each point.

Key Contributions

XiYOLO Framework: A novel approach to energy-adaptive perception that formulates the problem as designing a scalable base detector rather than searching for isolated models.
XiResOFA Search Space: A block design with controllable compression, kernel size, and attention that enables interpretable accuracy–energy tradeoffs.
Iterative Search Strategy: A progressive search method over backbone, FPN, and PAN that identifies a high-quality base architecture more efficiently than global joint search.
Two-Stage Energy Estimator: A model that combines a generic architecture predictor with a hardware-specific residual, significantly improving sample efficiency for energy-aware search under sparse hardware measurements.

Experimental Results

Experiments were conducted on the PascalVOC and COCO datasets, with real-device deployment on a ModalAI Sentinel Development Drone (featuring a Qualcomm Adreno 650 GPU and 15 TOPS NPU).

PascalVOC Performance: The medium XiYOLO model achieved 86.15 mAP50, surpassing YOLOv12m by +0.07 mAP50. Crucially, it reduced energy consumption by 20.6% on GPU and 35.9% on NPU compared to YOLOv12m. At the small scale, it achieved 84.38 mAP50 with energy reductions of 47.0% (GPU) and 45.7% (NPU).
COCO Performance: The medium model maintained competitive accuracy (50.71 mAP50-95) while reducing energy by 24.5% on GPU and 38.8% on NPU relative to YOLOv12m. At the small scale, energy was reduced by up to 53.7% on GPU and 51.6% on NPU.
Deployment Efficiency: Cumulative energy benchmarks over time showed that XiYOLO variants consistently accumulated less energy than YOLO baselines on both GPU and NPU, with the medium scale showing the most significant advantages.
Estimator Efficiency: On HW-NAS-Bench, the two-stage estimator achieved lower test RMSE than a joint predictor across 2–20 few-shot samples, demonstrating superior sample efficiency in adapting to new hardware.
Search Space Analysis: Randomly sampled models from the XiResOFA space showed an average energy reduction of 21.1% compared to the YOLOv12 baseline. The iterative search space yielded an average predicted mAP50 improvement of ~3.5% over a global search space.

Significance and Claims

The paper claims that XiYOLO successfully treats edge detector design as a search for budget-adaptive operating points aligned with heterogeneous hardware constraints, rather than a search for a single efficient model. By combining an interpretable search space with a sparse-data energy estimator, the framework achieves a stronger accuracy–energy tradeoff than existing YOLO baselines.

The authors emphasize that their approach enables interpretable accuracy–energy tradeoffs under sparse hardware measurements, making energy-aware design transferable across platforms where accurate energy data is scarce. The results suggest that searching for a single base architecture and scaling it is a viable strategy for creating detector families that perform well across diverse edge devices, extending system endurance for applications like search and rescue and infrastructure inspection without sacrificing detection quality.

The paper acknowledges limitations, noting that the search is confined to a YOLOv12-derived family, the final models are obtained via post-search scaling rather than multi-point search, and evaluation is limited to specific GPU and NPU settings. Nevertheless, the work demonstrates that energy-aware architecture search can produce detector designs that are both accurate and energy-efficient for heterogeneous edge deployment.