PaperlayerView exact revision on arXiv

Source-linked AI summary

WirelessGPT: A Generative Pre-trained Multi-task Learning Framework for Wireless Communication

Tingting Yang, Ping Zhang, Mengfan Zheng, Yuxuan Shi, Liwen Jing, Jianbo Huang, Nan Li

arXiv:2502.06877v1cs.LG

TL;DR

WirelessGPT addresses the need for unified multi-task wireless models by pretraining on large-scale channel data to learn representations usable across communication and sensing tasks. Its universal representations support downstream performance, including 98.11% testing recognition accuracy for human activity recognition and accurate environmental reconstruction.

  • Problem

    Wireless systems increasingly require unified support for channel modeling, signal prediction, environmental reconstruction, and target tracking, while conventional task-specific models lack flexibility and require substantial computational resources.

  • Method

    WirelessGPT uses unsupervised pretraining on Traciverse, SionnaRT, and DeepMIMO wireless channel datasets to learn universal representations capturing spatial, temporal, frequency, and multi-domain features for communication and sensing tasks.

  • Results

    98.11% testing recognition accuracy was achieved for human activity recognition using the universal representation, while environmental reconstruction showed high-accuracy learning of spatial and spectral relationships.

  • Takeaways & Limitations

    The unified representation supports communication and sensing downstream tasks with minimal fine-tuning, providing a scalable foundation for integrated sensing and communication systems.

Abstract

from arXiv · show

This paper introduces WirelessGPT, a pioneering foundation model specifically designed for multi-task learning in wireless communication and sensing. Specifically, WirelessGPT leverages large-scale wireless channel datasets for unsupervised pretraining and extracting universal channel representations, which captures complex spatiotemporal dependencies. In fact,this task-agnostic design adapts WirelessGPT seamlessly to a wide range of downstream tasks, using a unified representation with minimal fine-tuning. By unifying communication and sensing functionalities, WirelessGPT addresses the limitations of task-specific models, offering a scalable and efficient solution for integrated sensing and communication (ISAC). With an initial parameter size of around 80 million, WirelessGPT demonstrates significant improvements over conventional methods and smaller AI models, reducing reliance on large-scale labeled data. As the first foundation model capable of supporting diverse tasks across different domains, WirelessGPT establishes a new benchmark, paving the way for future advancements in multi-task wireless systems.

I. INTRODUCTION

WirelessGPT is proposed as a unified foundation model for wireless communication and sensing, addressing task specialization, weak cross-domain generalization, and limited labeled data. Its design combines multidimensional pretraining, scalable representations, and the Traciverse dataset for diverse downstream tasks.

  • Motivation and proposal: Conventional and recent learnable models remain task-specific, generalize poorly across spatial, temporal, and frequency domains, and often require substantial labeled data.These limitations create a bottleneck for unified ISAC systems.
  • Motivation and proposal: WirelessGPT targets multi-task wireless communication and sensing within a unified framework.It is intended to support integrated communication and sensing applications with minimal fine-tuning.
  • Key contributions: WirelessGPT’s pretraining captures spatial, temporal, and frequency correlations in wireless channel data for comprehensive representation learning.The approach is designed to exploit multidimensional channel structure across diverse communication tasks.
  • Key contributions: Pretrained channel representations support channel prediction, environmental reconstruction, and localization while extending the model to sensing functionalities.The representations use cross-dimensional correlations for downstream execution.
  • Key contributions: 80M parameters is the initial model size, with a stated path toward 800M parameters for increasingly complex communication and sensing tasks.The contribution frames model size as scalable rather than fixed.
  • Key contributions: Traciverse is a publicly available wireless channel dataset spanning over 27 cities and more than 100 diverse scenarios.The dataset is approximately 300 GB and covers environments including dense urban, residential, urban-center, and suburban settings.

II. LITERATURE REVIEW

The literature review describes progress toward wireless foundation models but emphasizes that existing approaches remain limited by task focus or incomplete cross-domain unification. WirelessGPT is positioned as a scalable foundation model for diverse ISAC communication and sensing tasks.

  • Research gap: Traditional wireless approaches struggle to generalize across diverse environments and often require large amounts of labeled data.These limitations motivate unified, task-agnostic alternatives.
  • Existing approaches: Prior work includes language-model-assisted physical-layer communication, but the cited framework is limited to single-task scenarios.The review contrasts single-task robustness with broader multi-task requirements.
  • Existing approaches: The Large Wireless Model applies Transformer-based self-supervised pretraining to multiple tasks such as channel modeling and beamforming, while focusing primarily on physical-layer functions.The review presents it as related task-agnostic work rather than a complete communication-sensing framework.
  • Existing approaches: Other studies explore contrastive multimodal learning, masked spectrogram modeling, digital-twin channel generation, and sensor fusion, but address narrower task or modality combinations.The cited examples include zero-shot learning, radio foundation models, channel generation, and beam prediction.
  • Existing approaches: Self-supervised localization and transformer-masked autoencoder studies demonstrate downstream potential but remain focused on specific tasks rather than comprehensive multi-task learning.The review specifically notes limitations spanning localization, channel estimation, and resource allocation.
  • Positioning WirelessGPT: WirelessGPT builds on these developments by pretraining on large-scale wireless channel datasets and fine-tuning across diverse communication and sensing tasks in ISAC.The paper presents this as a scalable and unified solution to existing limitations.

III. FRAMEWORK

WirelessGPT is pretrained using multiple wireless-channel datasets to generate effective channel representations. The framework combines the self-developed Traciverse dataset with SionnaRT and DeepMIMO, whose simulation pipelines provide varied propagation and scenario characteristics.

  • Framework overview: WirelessGPT’s framework includes data preprocessing, cross-domain representations, positional encoding, and optional fine-tuning.The model is described as a Transformer-based foundation model for wireless communication channels.
  • Pretraining data: Three datasets—Traciverse, SionnaRT, and DeepMIMO—are incorporated into pretraining to generate channel representations for downstream tasks.The stated goal is to improve downstream performance through representation learning.
  • Pretraining data: Traciverse contains 300 GB of publicly available data from over 27 cities and more than 100 scenarios across multiple 6G candidate frequency bands.It is generated with NVIDIA Sionna and captures propagation phenomena including reflection, refraction, and scattering.
  • Pretraining data: SionnaRT uses ray tracing and high-fidelity scene maps from Sensiverse to generate multipath channels across 25 urban and rural environments.The framework captures reflection, scattering, and diffraction for realistic simulations.
  • Pretraining data: DeepMIMO uses ray-tracing data from Wireless InSite and supports customization of scenarios through system and channel parameters.Its propagation characteristics reflect environmental geometry, materials, and transmitter-receiver configurations.

B. Model Pretraining

WirelessGPT preprocesses wireless signals as three-dimensional time-space-frequency slices and encodes them with a multi-domain Transformer. Cross-domain self-attention produces joint representations, while model scale can be adjusted for different task complexities and resources.

  • Input representation: Wireless signals are organized into uniform three-dimensional slices spanning time, space, and frequency before model input.Snapshots represent time, antenna geometry represents space, and subcarriers represent frequency.
  • Transformer encoding: A multi-domain Transformer encoder extracts features using patch-based spatial embeddings, causal temporal self-attention, and frequency-domain representations.These components preserve geometric structure and model temporal variation across wireless data.
  • Transformer encoding: Cross-domain self-attention integrates the domain embeddings into joint representations of cross-domain dependencies.Positional encodings are applied across domains except the temporal domain, where they are described as unnecessary.
  • Scalability: WirelessGPT supports model sizes from around 600K to 800M parameters for tasks with different complexity and computational-resource requirements.Depth, attention heads, and hidden units can be adjusted to balance computational efficiency and capacity.

4) Pretraining Process:

WirelessGPT is pretrained through masked reconstruction on large-scale wireless datasets to learn universal channel representations. These representations are then generated from downstream-task data and supplied to task-specific models, with optional fine-tuning for especially accuracy-sensitive tasks.

  • Masked-patch reconstruction during unsupervised pretraining uses contextual information to learn meaningful representations and cross-domain dependencies.
  • The structured pipeline combines pretraining, representation generation, and downstream-task execution.
  • The application pipeline begins with extensive pretraining on large-scale wireless datasets to learn universal wireless-channel representations.
  • For each downstream task, the pretrained model converts task data into a universal representation that captures essential signal features.
  • The generated representation is fed to a downstream model to improve task accuracy and robustness.

4) Fine-Tuning (Optional):

WirelessGPT supports downstream communication and sensing tasks through universal representations, with optional task-specific fine-tuning for regression tasks. In channel estimation, it is evaluated against baseline models using NMSE and complexity measurements, while human activity recognition demonstrates fast convergence and high accuracy.

  • 4) Fine-Tuning (Optional):: Fine-tuning is optional for exceptionally high-accuracy tasks and is applied to regression tasks but omitted for classification tasks.
  • Downstream evaluation: WirelessGPT is adapted to channel estimation, channel prediction, and human activity recognition to assess generalization across communication and sensing tasks.
  • Channel estimation: Channel estimation uses an OFDM system with 32 sub-carriers, a 32-antenna array, a 64-length pilot, QPSK modulation, and average SNR from -5 to 10.
  • Channel estimation: NMSE is used to quantify channel-estimation error when comparing universal representations and raw channel data with downstream Transformer and ResCNN models.
  • Complexity: WirelessGPT’s channel representations reduce downstream training and inference time by converting complex channel matrices into more structured inputs.

B. Channel Prediction

Channel prediction forecasts future channel states from historical behavior and is evaluated against Transformer, LSTM, and LLM-based baselines. WirelessGPT’s representation improves sequential prediction, matches LLM4CP in some conditions, and offers lower time complexity than smaller models, though the LLM4CP timing comparison uses different GPUs.

  • Channel prediction forecasts future wireless-channel states from estimated historical behavior.
  • Experimental setting: The experiment uses a MISO-OFDM system with a 4 × 4 antenna array, QuaDRiGa-generated channels, a 2.4 GHz center frequency, and 32 subcarriers.
  • WirelessGPT versus Transformer/LSTM: NMSE is plotted against SNR at velocities of 40 and 90 km/h for WirelessGPT, Transformer, and LSTM competitors.
  • WirelessGPT versus LLM4CP: WirelessGPT with a 1.7M-parameter Transformer has comparable NMSE to LLM4CP, outperforming it below 15 SNR but underperforming in pure channel conditions.
  • Complexity: WirelessGPT’s efficient channel representation improves training and inference time complexity even relative to smaller models.
  • Complexity: The LLM4CP training and inference times are not directly aligned because the comparison uses different GPU devices.

C. Human Activity Recognition

The human activity recognition case study uses WiFi CSI amplitude measurements and a universal representation generated by WirelessGPT for compact downstream classification. The representation achieves high recognition accuracy despite the measurement data differing from the pretraining data.

  • Dataset and task: The task classifies six human activities from WiFi CSI amplitude measurements collected by three routers from 20 subjects.The activities are running, walking, falling down, boxing, circling arms, and cleaning the floor.
  • Representation generation: WirelessGPT samples and reshapes each measurement before its encoder generates a compact universal representation retaining essential input features.The representation is used as input to the downstream classifier rather than the raw measurements.
  • Results: 98.11% testing recognition accuracy is achieved with a residual 2D CNN classifier using 316.1K parameters.Training loss converges within 3 minutes on a GeForce 3060 GPU, and the result is reported as comparable to state-of-the-art performance.
  • Generalization: The evaluation data contains indoor WiFi CSI amplitude measurements, while pretraining used outdoor complex-valued CSI matrices between base stations and end users.This cross-setting evaluation is presented as evidence of the foundation model’s generality.
  • Compression: The universal representation reduces the downstream classifier input to 0.67% of the original measurement data.The compressed representation preserves the input features while providing denser information.

D. Wireless Reconstruction

Wireless reconstruction is evaluated as part of the convergence of wireless communication and environmental sensing in ISAC systems. The experiments cover LOS and NLOS propagation across urban and rural environments and frequencies from 3.5 GHz toward millimeter-wave bands.

  • ISAC motivation: ISAC systems integrate wireless communication with environmental sensing for applications including positioning, environmental monitoring, and autonomous navigation.The section frames wireless-environment reconstruction as relevant to this integrated functionality.
  • Evaluation conditions: Environmental reconstruction is evaluated under both line-of-sight and non-line-of-sight conditions, including first-order reflections.The NLOS setting uses Sionna’s MaxNumReflections parameter set to 1.
  • Evaluation conditions: Experiments span urban and rural environments and frequencies from 3.5 GHz to 10 GHz and 28 GHz millimeter-wave bands.The frequency range is used to test reconstruction across low-frequency and higher-frequency settings.

1) Results:

The environmental reconstruction experiments assess whether WirelessGPT can learn spatial and spectral relationships from multimodal inputs. Training behavior is examined through loss convergence using Chamfer Distance.

  • Training results: The environmental reconstruction training loss converges consistently, particularly the Chamfer Distance loss.This convergence is reported as indicating effective learning of spatial and spectral relationships in the multimodal input data.
  • Training strategy: The evaluation uses mini-batch processing, positional encodings, and Chamfer Distance as the reconstruction loss metric.These components form the training strategy described for the reconstruction experiments.
  • Simulation configuration: The simulated setup uses one omnidirectional transmitting antenna and 16 omnidirectional receiving antennas arranged in a 4 × 4 planar array.The receiving-array spacing is set to half the wavelength at a 10 GHz carrier frequency.

2) Analysis:

WirelessGPT is reported to reconstruct wireless environments accurately across diverse scenarios and frequency bands. The paper also identifies larger-scale models and broader applicability as directions for future evaluation.

  • Analysis: WirelessGPT achieves high-accuracy wireless-environment reconstruction with adaptability across diverse scenarios.The reported evaluation emphasizes performance across complex propagation and dynamic environments.
  • Analysis: The model achieves superior resolution in both low-frequency and millimeter-wave bands while capturing strong scattering features.The paper attributes this capability to the generative capabilities of its Transformer architecture.
  • Future work: Future work will scale WirelessGPT to larger parameter sizes to examine scaling laws and broader applicability.The paper presents this as part of establishing a single model for communication and sensing tasks.
Loading 2502.06877v1…