The rapid growth in the parameter size of large language models (LLMs) has introduced significant challenges for deployment on edge devices. To address these challenges, this paper focuses on developing a post-training quantization (PTQ) framework tailored for LLM deployment on edge devices. Our framework introduces an enhanced channel smoothing technique based on channel value ranges, combined with channel reordering to mitigate quantization errors associated with large channel value range differences and activation outliers, reducing the need for quantization and dequantization steps during inference, making it more suitable for edge deployment. Our approach achieves full integer quantization for all model operations, reducing model size by 4×. Through extensive experiments on various language tasks using the OPT model, we demonstrate that our framework surpasses state-of-the-art methods under the W4A4 configuration.

Advanced driver assistance systems (ADASs) enhance driving safety and convenience by providing auxiliary functions. However, traditional rule-based or learning-based ADAS lack the capability for commonsense-based environmental understanding and multisensor data fusion, which leads to limitations in complex dynamic environments. Multimodal large language models (MLLMs) can effectively integrate data from different modalities and possess strong environmental perception and commonsense reasoning abilities, offering more intelligent driver assistance services within Internet of Things (IoT) networks. In this article, we propose a cloud-edge collaborative ADAS based on MLLMs, utilizing IoT networks by deploying a smaller model, CogVLM2, at the edge and a larger model, ChatGPT-4o, in the cloud to achieve collaborative driver assistance services. Specifically, we first reannotate the BDD-X dataset and use it to fine-tune CogVLM2 with LoRA, while applying few-shot learning to ChatGPT-4o to enhance their understanding and decision-making capabilities in traffic scenarios. We then formulate service latency, energy consumption, and Quality-of-Service (QoS) models for the cloud-edge collaborative ADAS in IoT networks, optimizing the combination of these models. Finally, we design an improved DDPG-based task offloading algorithm by introducing a multistep reward mechanism and using a diffusion model to generate noise, aiming to determine the optimal execution location (i.e., cloud, edge, or local) for each task. Experimental results show that both CogVLM2 and ChatGPT-4o can achieve basic ADAS functionality. After fine-tuning and few-shot learning, their task success rates were significantly improved. Moreover, compared to other mainstream deep reinforcement learning-based task offloading algorithms, the improved DDPG task offloading algorithm demonstrates better performance in latency, energy consumption, and QoS within IoT networks.

The growing focus on indoor robot navigation utilizing wireless signals has stemmed from the capability of these signals to capture high-resolution angular and temporal measurements. Prior heuristic-based methods, based on radio frequency (RF) propagation, are intuitive and generalizable across simple scenarios, yet fail to navigate in complex environments. On the other hand, end-to-end (e2e) deep reinforcement learning (RL) can explore a rich class of policies, delivering surprising performance when facing complex wireless environments. However, the price to pay is the astronomical amount of training samples, and the resulting policy, without fine-tuning (zero-shot), is unable to navigate efficiently in new scenarios unseen in the training phase. To equip the navigation agent with sample-efficient learning and zero-shot generalization, this work proposes a novel physics-informed RL (PIRL) where a distance-to-target-based cost (standard in e2e) is augmented with physics-informed reward shaping. The key intuition is that wireless environments vary, but physics laws persist. After learning to utilize the physics information, the agent can transfer this knowledge across different tasks and navigate in an unknown environment without fine-tuning. The proposed PIRL is evaluated using a wireless digital twin (WDT) built upon simulations of a large class of indoor environments from the AI Habitat dataset augmented with electromagnetic radiation simulation for wireless signals. It is shown that the PIRL significantly outperforms both e2e RL and heuristic-based solutions in terms of generalization and performance. Source code is available at https://github.com/Panshark/PIRL-WIN.

High-rank line-of-sight (LOS) MIMO systems have attracted considerable attention for millimeter wave and THz communications. The small wavelengths in these frequencies enable spatial multiplexing with massive data rates at long distances. Such systems are also being considered for multi-path non-LOS (NLOS) environments. In these scenarios, standard channel models based on plane waves cannot capture the curvature of each wave front necessary to model spatial multiplexing. This work presents a novel and simple multi-path wireless channel parametrization where each path is replaced by a LOS path with a reflected image source. The model is fully valid for all paths with specular planar reflections, and captures the spherical nature of each wave front. Importantly, it is shown that the model uses only two additional parameters relative to the standard plane wave model. Moreover, the parameters can be easily captured in standard ray tracing. The accuracy of the approach is demonstrated on detailed ray tracing simulations at 28GHz and 140GHz in a dense urban area.

In this paper, we study a navigation problem where a mobile robot needs to locate a mmWave wireless signal. Using the directionality properties of the signal, we propose an estimation and path planning algorithm that can efficiently navigate in cluttered indoor environments. We formulate Extended Kalman filters for emitter location estimation in cases where the signal is received in line-of-sight or after reflections. We then propose to plan motion trajectories based on belief-space dynamics in order to minimize the uncertainty of the position estimates. The associated non-linear optimization problem is solved by a state-of-the-art constrained iLQR solver. In particular, we propose a method that can handle a large number of obstacles (∼ 300) with reasonable computation times. We validate the approach in an extensive set of simulations. We show that our estimators can help increase navigation success rate and that planning to reduce estimation uncertainty can improve the overall task completion speed.