Can the Cloud Drive? Infrastructure Feasibility of Offloading Autonomous Driving Across 5G and 6G

Making the Most of Predictions: Data Offloading for Autonomous Driving

Autonomous driving struggles to anticipate events due to the limitations of first-person perception. By offloading sensor data, environmental awareness can be shared between road users and greatly improve perception of obstacles, leading to safer and more versatile trips. To this end, vehicular networks require stable, high-performance connectivity across heterogeneous Radio Access Technologies (RAT), yet existing selection mechanisms react to degradation rather than anticipating it. We present a predictive quality-of-service (QoS) framework that jointly forecasts latency and packet delivery rate using dual-output recurrent neural networks, enabling proactive RAT selection across 5G NR, C-V2X, and DSRC. Where most frameworks tend to stop at link continuity, our proposal operates as a hierarchical closed-loop system that also maximizes throughput: a model-based selector chooses the optimal RAT, then a discrete-time Markov chain adapts packet size to channel conditions, then a queue manager enforces per-RAT capacity constraints, leading to new QoS measurements and online retraining, closing the loop. To study scalability, we introduce per-RAT contention models grounded in each RAT’s scheduling standard, and evaluate a two-pass contention-aware selection policy under increasing population sizes. Field trial data collected on an urban platform using V2X hardware across three simultaneous RATs validates the approach. The predictive QoS scheme achieves 75.2% of transmissions above 99% PDR versus 63.7% for the reactive baseline, with 40–60% fewer RAT handovers. Upscaled simulations show that contention-aware selection yields measurable reliability gains at medium and high densities, where shared-medium RATs face meaningful resource pressure. Dynamic packet sizing further improves per-vehicle throughput by up to 12.6% by exploiting high-reliability windows to increase payload.

Risk-Adaptive Edge--Cloud Visual Reasoning for Communication-Efficient Autonomous Driving

Cloud-hosted vision-language models (VLMs) offer greater contextual reasoning capabilities than smaller onboard models, but frequent visual uploads increase communication overhead and add network and inference latency to tactical decisions. We present a risk-adaptive edge-cloud architecture in which onboard traffic assessment determines when cloud reasoning is requested. An onboard VLM and a lightweight detector capture temporal traffic conditions and path-relative hazards for conservative local response and selective cloud access. The cloud model provides tactical advice, while validation, vehicle control, and automatic emergency braking remain local. In CARLA experiments, our method matched the task success rate of periodic cloud access while reducing cloud requests by 54.1% and recording fewer automatic emergency braking (AEB) activations. In a delayed-roadwork ablation, semantic events triggered requests before the next scheduled audit. Across three emulated network profiles, the method continued to reduce cloud traffic, although lane changes took longer than with periodic access. Onboard traffic assessment therefore served as a practical trigger for selective VLM inference in these experiments.

Chasing Clouds with Donkeycar: Holistic Exploration of Edge and Cloud Inferencing Trade-Offs in E2E Self-Driving Cars

TALSC: Timeliness-Aware Large-Small VLM Collaboration for Infrastructure-Assisted Autonomous Driving

The deployment of Vision-Language Models (VLMs) in autonomous driving (AD) systems is constrained by on-board computing power, restricting vehicles to small VLMs (SVLMs) with limited perception and reasoning capabilities. Infrastructure-assisted AD alleviates this resource constraint by enabling collaboration with large VLMs (LVLMs) at edge servers. However, in dynamic vehicular environments, the utility of sensory data for downstream tasks decays rapidly, making timeliness of information a critical concern. To balance the accuracy gains of LVLMs with their latency-induced timeliness degradation, we develop a Timeliness-Aware Large-Small VLM Collaboration (TALSC) framework. Specifically, we first model the Age of Information (AoI) evolution for VLM inference and characterize the coupling among AoI, token length, and task performance to formulate a general timeliness metric. Building on this, we propose the TALSC online scheduling algorithm. Since scheduling decisions have a delayed impact on future timeliness metric and the output token number is unknown at scheduling time, we design a Lyapunov drift-plus-estimated-penalty algorithm and provides a guaranteed performance. In simulation, we first conduct a case study to derive a fitted timeliness metric based on nuScenes dataset, and further show that TALSC outperforms baselines under various communication and computing settings, achieving up to a 12.6\% normalized improvement in Micro-F1 score compared with the best-performing baseline.

Energy-Optimized Lightweight DRL for Computational Offloading in UAV-IoV Systems

The integration of 5G/6G networks with the Internet of Vehicles (IoV) requires efficient computational offloading for data-intensive applications such as autonomous driving and augmented reality. Although Unmanned Aerial Vehicles (UAVs) offer agile mobile edge computing (MEC) capabilities, their operational efficiency is hampered by high mobility, limited battery life, and the complexity of joint resource optimization. Existing offloading strategies often fail to simultaneously optimize latency, energy consumption, and resource utilization under dynamic IoV conditions. This paper proposes a novel Energy-Optimized Lightweight Deep Reinforcement Learning (DRL) framework for intelligent task offloading in UAV-assisted IoV networks. Our approach leverages a simplified Double Deep Q-Network (DDQN) to dynamically manage task partitioning by intelligent offloading decisions, UAV trajectory planning through optimized path forecasting, and resource allocation through adaptive computation distribution. Key innovations include a streamlined state-space design that reduces computational overhead by 30% and a composite reward function that balances latency and energy objectives. These are realized by a prioritized experience replay mechanism and a target network separation strategy that enhances learning stability. Experimental results demonstrate that our framework achieves a task success rate of 98.5%, reduces latency by 40%, and maintains a 78.1%. The results confirm the framework’s superiority, demonstrating significant improvements over its base architecture (DQN), its enhanced variant (DDQN), and other state-of-the-art baselines like MADDPG and game-theoretic approaches, thereby providing a robust solution for practical UAV-IoV deployments.

From Intent to Infrastructure: LLM-Driven Agent Compilers for ISAC Networks

Integrated sensing and communications (ISAC) is moving from proof-of-concept demonstrations to system-level deployment in sixth-generation (6G) networks. Because sensing and communication share hardware, spectrum, and waveform resources, ISAC design now involves many tightly coupled choices, including waveform selection, sensing algorithm setup, resource scheduling, and deployment planning. This design space is already too large to manage well through manual tuning or isolated optimizers. This article introduces the \textit{Agent Compiler}, a large language model (LLM)-enabled compilation layer that translates high-level engineering intent into complete and executable ISAC system configurations. The Agent Compiler works in four stages: intent parsing, task decomposition, policy graph synthesis, and infrastructure mapping. It produces a verifiable intermediate representation called the ISAC Policy Graph (IPG). A runtime engine then deploys the compiled configuration and supports closed-loop adaptation at three levels: fast parameter updates, partial recompilation of affected subgraphs, and full workflow recompilation. The core design principle is strict time-scale separation: the LLM handles slow-loop strategic decisions, while proven algorithms retain real-time control in the fast loop. A UAV-assisted disaster rescue example illustrates the full compilation process. We also discuss open issues, including compilation latency, output reliability, constraint verification, and pipeline security, to guide future research.