Energy-Efficient adaptive perception for autonomous driving via lightweight policy learning and simulation-based optimization

Abstract

Modern autonomous driving systems rely heavily on deep learning-based perception models for object detection; yet, their computational and energy demands remain critical bottlenecks. The existing adaptive-perception strategies often lack the ability to dynamically balance the detection accuracy and energy consumption, in real-time, particularly under varying environmental conditions. To address this challenge, we first construct a large-scale autonomous driving dataset based on the CARLA simulator. Then, we propose a novel metric—the balanced efficiency index—to annotate each image with the most suitable you-only-look-once version 8 (YOLOv8) model size (i.e., n, s, m, l, or x). This index is governed by two critical parameters, which are efficiently optimized using our proposed constrained stochastic DIviding RECTangles (DIRECT) algorithm. Finally, we propose a lightweight dynamic mixed receptive field transformer (DynaMixFormer), which is trained using the labelled dataset, to select the appropriate YOLOv8 model adaptively. Our results show that: (1) the constrained stochastic DIRECT algorithm determines cost-effective parameters with very limited simulation overhead; (2) DynaMixFormer achieves a high classification accuracy of 96.56 % with only 0.017 M parameters, outperforming the state-of-the-art image-classification networks; and (3) the well-trained DynaMixFormer effectively extracts real-time contextual features, such as traffic density, weather conditions, and road complexity, to intelligently select the optimal model from various YOLOv8 variants. Extensive simulations demonstrate that our approach achieves up to 70.20 % reduction in the energy consumption, compared to the static deployment of the YOLOv8x model, with only a marginal decrease of approximately 2 % in the mean average precision. Taking China as an example, this translates to an estimated energy saving of 2.73 × 10¹⁴ W. This work not only advances energy-efficient autonomous perception but also provides a generalizable framework for adaptive model selection in resource-constrained edge-computing systems. For ease of comprehension, some key nomenclature used in this paper are summarized in Table 1.