Transformer-Based Task-Oriented Joint Source-Channel Coding for Edge AI Agent

In the sixth-generation (6G) mobile networks, a large number of artificial intelligence (AI) agents, such as embodied intelligence systems and autonomous vehicles, will require real-time decision-making based on high-throughput, low-latency sensor data streams. Task-oriented Joint Source-Channel Coding (JSCC) has emerged as an effective paradigm by aligning communication strategies with AI tasks. In this paper, we introduce a Transformer-based task-oriented JSCC framework tailored for image transmission in autonomous driving scenarios. The proposed architecture integrates a MobileNetinspired feature extractor with Swin-Transformer blocks in the encoder to capture features for the AI task, while the decoder adopts a symmetric design with PixelShuffle-based upsampling. We evaluate the proposed architecture under varying channel conditions in an edge intelligent system for AI driving agents. The results show that our model significantly outperforms the existing baselines, improving driving score performance by 35.6%, while reducing floating-point operations and the model size by 70.9% and 35.4%, respectively. Additionally, our method substantially enhances perceptual image quality, surpassing the baseline by 119.57% in Peak Signal-to-Noise Ratio (PSNR), 467.21% in Multi-Scale Structural Similarity Index (MS-SSIM), and 43.37% in Learned Perceptual Image Patch Similarity (LPIPS).