GTI & GSMA Challenge 2026: AI Evolution for Communication and Sensing in Future Network

Background

The U6G multi-station collaborative sensing system overcomes the physical limitations of single-station sensing through distributed node coordination. Its high-precision positioning, occlusion-resistant imaging, and environmental reconstruction capabilities provide a novel approach for ground target recognition.

Barrier

However, when applied to complex ground scenarios, the technology still faces the core challenge of severe clutter interference. Static clutter (buildings, vegetation) and dynamic clutter (vehicles, pedestrians) intertwine, resulting in low signal-to-noise ratios and obscured target features, which hinders effective target detection.

The inherent defects of traditional perception model, such as insufficient generalization ability and lack of prior knowledge in complex electromagnetic environment, need to be solved from the bottom level of perception logic.

Key Areas of Focus

AI models provide innovative approaches for clutter elimination. By integrating electromagnetic mapping and pre-training multi-scenario electromagnetic physical laws with channel data, they can acquire prior knowledge of electromagnetic propagation and scattering, thereby constructing a digital twin electromagnetic space that dynamically synchronizes with the real world and maintains high-fidelity mapping. The “data-driven + physical law fusion” model effectively addresses the core bottleneck in U6G multi-station collaborative sensing.

By introducing a dedicated electromagnetic model with its unique environmental understanding, we establish it as the core engine for clutter feature extraction and environmental modeling. This specialized model forms the basis of a knowledge-enhanced clutter suppression framework. Through in-depth analysis of raw sensing signals from multiple stations, the system accurately decodes electromagnetic scattering components from both static scenes and moving targets, enabling precise separation of clutter and target signals. This approach overcomes environmental modeling constraints, enhances clutter suppression robustness in complex scenarios, and ultimately improves ground target localization accuracy.

Goal of the Challenge – dedicated model-based ground environmental clutter cancellation

This Challenge focuses on dedicated model-based ground environmental clutter cancellation technology for collaborative sensing scenarios.

1) The multi-station sensing system consists of three components: the signal transmission node A, and the sensing nodes B and C.

2) Model input data: frequency-domain data from the receiving antennas of sensing nodes B and C.

3) Model outputs data：delay, Doppler and angle information of the perceived target.

4) Model functions: It features target recognition via sensing receivers and dynamic-static clutter cancellation capabilities.

5) Supports maximum configuration: 64 antennas, 400MHz bandwidth, and 120KHz subcarrier spacing.

6) Performance gains: Compared to traditional sensing receivers, the system achieves sub-meter-level positioning accuracy after clutter cancellation, with false alarm and missed detection rates below 1%.

Specific Devices and Capabilities for Task 1

For teams participating in Challenge Task 1, we provide a dataset , which serves as the foundational data for model training.

1. Data Set

The dataset serves as the foundation for training and validating clutter removal models. It must encompass three core components: multi-station collaborative sensing, electromagnetic models, and ground environmental clutter, ensuring the data’s authenticity, diversity, and annotatability.

Multi-Station Synergy: Synchronized data collection from three or more base stations, supporting multi-scenario collaboration with multi-perspective clutter suppression and fusion perception. Electromagnetic Authenticity: The output integrates electromagnetic models, with clutter types and parameters (e.g., SNR, time delay spread, Doppler shift) labeled to ensure data reflects real-world electromagnetic environments. Scenario Diversity: Covers urban macro station scenarios with both stationary and moving targets to ensure model generalization capability.

A multi-site collaborative sensing network, composed of multiple base stations (nodes) operating on U6G frequency bands with 400MHz bandwidth and 120kHz subcarriers. The data content is detailed in the table below.

No.	Data content	acquisition end	Delivery method	description
1	B node frequency domain data	B-side of the base station	TXT files	Supporting multi-slot multi-antenna data at node B
2	C node frequency domain data	C-side of the base station	TXT files	Supporting multi-slot multi-antenna C-node data
3	system configuration	A, B, C – on the base station side	doc files	A, B, and C station information, and the configuration parameters of the sensing reference signal.
4	target information	Environmental side	TXT files	includes ground static clutter, dynamic clutter, and target reflections

For teams participating in the challenge, the model must undergo systematic validation using test datasets. The evaluation will focus on core metrics including positioning accuracy, false alarm rate, and missed detection rate to identify the most comprehensive dedicated model. The winning electromagnetic-specific model will be documented with corresponding code. The device also offers two integration options: access to an open high-fidelity simulation platform or a hardware device platform. In real-world air interface environments, the model’s engineering performance will be validated to establish a closed-loop optimization solution integrating data, model, and device.

2) Interface Method 1: Open High-Fidelity Simulation Platform Integration

The open high-fidelity simulation-platform establishes a hardware-agnostic mirror environment for base station communication and sensing functions. Its sensing module integrates core capabilities including synchronization, parameter estimation, target detection, and clutter cancellation. Through software decoupling and flexible customization interfaces, the platform supports external model integration, creating a highly realistic simulation environment for technologies like U6G clutter cancellation. This accelerates algorithm iteration and deployment.

Integration method: Provides module-level/system-level communication simulation and implementation code, enabling modular replacement (e.g., clutter cancellation, target detection modules). The modular software architecture allows users to integrate custom algorithm modules without system-wide refactoring.

Interface file ：Matlab file, which is portable and can be compiled and run.

Input interface: Frequency-domain data, receiving spatial IQ signals from multi-station cooperative sensing while preserving original electromagnetic characteristics.

Output interface: Standardized core parameters for target perception, including target delay, Doppler shift, and angle information (structured data), directly supporting applications such as positioning and tracking.

Algorithm Deployment: The system adopts a hot-swapping approach using simulation code to replace the traditional clutter cancellation algorithm with the newly proposed algorithm. During the replacement, input and output are mapped through an interface adapter to ensure seamless integration with the original module, enabling rapid iterative verification.

3) Interface Method 2: Open Hardware Device Platform Integration

The hardware platform consists of BBU (Base Band Unit) and AAU (Active Antenna Unit). As the core processing node of the base station, the BBU features a well-defined architecture of CPU + FPGA + GPU, providing a low-latency, high-performance, and scalable hardware foundation for the large-scale electromagnetic model. Specifically:

The CPU handles task scheduling (including process and timing control), lightweight preprocessing, and base station system management, serving as the core of the entire system.

FPGA: Enables high-speed data processing and interface adaptation (e.g., fiber-optic communication with RRU), enhancing the system’s parallel computing capabilities.

GPU: GPU with 10,240 CUDA cores delivers high-precision inference for large-scale AI models. It supports multi-GPU parallel processing, meeting real-time processing requirements with 400MHz bandwidth and 120kHz subcarrier spacing, while achieving sub-200ms latency.

The model must interface with the BBU’s data path using the ONNX format as the standard, with the following procedure:

Deployment method: The trained model is deployed on the GPU.

Format: ONNX.

Verification scenario: Station A sends, Station B and Station C receive.

Input interface: The BBU’s FPGA transmits the IQ data (frequency-domain data after FFT transformation) sent by the AAU to the GPU via the PCIe bus, then passes it to the model through a CPU-GPU direct connection.

Processing workflow: The GPU loads the trained AI model to process IQ data, meeting the real-time requirements for target localization. Considering the floating-point operations in model training, the GPU incorporates fixed-point floating-point conversion processing for flexible integration.

Output interface: The GPU delivers operational results (including latency, Doppler, and angular accuracy) to the BBU’s CPU via the PCIe bus, which are then used for target recognition and tracking.

Goal

Ultimately, the “modules, algorithms, and models” scale replacement mechanism enables rapid technological iteration, accelerating the closed-loop process from simulation to implementation for algorithms like U6G clutter cancellation.

Background

Traditional base station receivers are built around fixed algorithms and dedicated hardware, relying on serial processing modules like LMMSE channel estimation, equalization, and demodulation. Their performance is based on the ideal assumptions of “stationary channels, Gaussian noise, and lossless hardware.”

Barrier

In the Future Network era, they face three core pain points:

1) Poor adaptability to complex environments. In practical scenarios, factors such as mobility, interference, and multipath fading can disrupt the model assumptions of traditional algorithms, leading to increased bit error rates and suboptimal performance.

2) Limited flexibility and adaptability. Traditional algorithms require manual pre-setting of parameters (e.g., equalizers, channel estimation) and cannot dynamically adjust to channel variations (e.g., PRB scheduling, MCS handover). When encountering new interference or bursty scenarios, reconfiguration is required, resulting in sluggish response.

3) Hardware dependencies and upgrade bottlenecks exist. Core modules (such as synchronization and channel estimation) are FPGA/ASIC fixed, resulting in lengthy development and upgrade cycles, making it difficult to support new features like Future Network ultra-large-scale MIMO.

Key Areas of Focus

As a core technology for Future Network and future networks, base station AI receivers focus on overcoming the systemic bottlenecks of traditional receivers in complex environments, specifically addressing the ‘performance-latency-generalization’ trade-off. Through data-driven approaches, these receivers integrate all components into an end-to-end learning model for joint optimization, enabling more accurate and stable signal reconstruction.

To ensure robust performance, the AI receiver must maintain consistent reception capabilities across diverse scenarios, including varying channel conditions, hardware configurations, and PRB/MCS scheduling schemes.

To achieve sufficient performance gain, the model must outperform traditional algorithms such as the LMMSE channel estimation and equalization receiver.

To ensure low inference latency, the model must maintain lightweight performance while preserving reception quality. This enables it to meet the high real-time processing demands of Future Network base stations—particularly edge nodes—when deployed, with Future Network 400MHz bandwidth and 120kHz subcarrier spacing.

Goal of the Challenge – developing an AI model for Future Network AI PUSCH receivers

This Challenge focuses on developing an AI model for Future Network AI PUSCH receivers.

Inference delay: Keep it below 1000 microseconds, with jitter not exceeding ±100 microseconds

Model input data: frequency domain data of the base station’s receiving antenna

Model output data: LLR soft bit data after receiver demodulation

Model features: includes channel estimation, equalization, and demodulation modules of a traditional receiver, but excludes decoding

Supports maximum configuration: 16 antennas, 400MHz bandwidth, 120KHz subcarrier spacing

Performance gain requirement: Compared to traditional receivers, especially those employing LMMSE channel estimation and LMMSE equalization, achieving the same block error rate (BLER) level requires a signal-to-noise ratio (SNR) gain of over 2 dB.

Lightweight requirement: fewer than 500,000 model parameters

Specific Devices and Capabilities for Task 2

For teams participating in Challenge Task 2, we provide a PUSCH dataset for AI receiver training.

The platform provides an open-source field measurement dataset for AI receiver model training, along with an open AI model inference infrastructure supporting model deployment and experimental testing. Specifically:

1. Dataset.

This dataset was based on U6G frequency band, 400MHz bandwidth, and 120KHz subcarriers.

The input content is listed in the table below.

No	Data content	acquisition end	Delivery method	description
1	PUSCH receives the antenna’s frequency-domain IQ data	Base station side	Txt, bin files	Supports 16-antenna data and both fixed-point and floating-point data
2	LLR Soft Bit Data of PUSCH Traditional Receiver	Base station side	Txt, bin files	Soft bits for multi-antenna data merging, supporting fixed and floating-point data
3	Channel measurement data	Base station side	Txt files	Real-time data collected by multiple antennas, such as antenna signal-to-noise ratio (SNR), time delay spread, and SNR range coverage exceeding 20dB
4	The bit-level data of the original transport block sent by PUSCH	Terminal side	Txt files	PUSCH supports pseudo-random sequence for each transmitted bit content
5	The bit-level data of the original transmission block sent by PUSCH after LDPC encoding	Terminal side	Txt files	The bit information of PUSCH after standard LDPC encoding
6	Community-level configuration information	Base station side	doc files	such as frequency points, bandwidth, physical cell ID, time slot configuration, number of receiving antennas, and number of scheduled users
7	User-level information	Terminal side	doc files	PUSCH scheduling information, including RB, MCS, DMRS, symbol configuration, and transmit antennas

Participant Requirements

For teams participating in the challenge, they must conduct systematic model validation using a dataset .

Evaluation Criteria

The evaluation will focus on core metrics such as the signal-to-noise ratio versus bit error rate (BER) curve, with the top-performing model being selected.

Expected Outcome

The winning model’s code will be collected and made available as an open-source AI inference framework for model deployment and experimental testing.

Further Equipment Details and Requirements

The Base Station Unit (BBU), serving as the core processing node of the base station, features a well-defined architecture of CPU + FPGA + GPU with clearly defined roles. This configuration provides the AI receiver model with a hardware foundation that ensures low latency, high computational power, and flexible scalability. Specifically:

The CPU handles receiver task scheduling (e.g., data distribution and process control), protocol processing, and BBU system management, serving as the system’s central processing unit.

FPGA: It handles high-speed data processing (e.g., encoding/decoding) and interface adaptation (e.g., fiber-optic communication with RRU), with its parallel computing capability compensating for CPU’s limitations in real-time performance.

GPU: As the AI inference core, it leverages NVIDIA’s architecture with 10,240 CUDA cores to deliver high-throughput, low-latency inference for AI receiver models. Through multi-card parallel processing, it meets real-time symbol-level requirements with 400MHz bandwidth and 120kHz subcarrier spacing.

The AI receiver model must interface with the BBU’s data path, with the following procedure:

Connection method: The trained AI receiver model is deployed on the GPU.

Format: ONNX.

Input interface: The BBU’s FPGA transmits the IQ data (frequency-domain data after FFT transformation) sent by the AAU to the GPU via the PCIe bus, while simultaneously delivering channel measurement data (e.g., signal-to-noise ratio, Doppler shift) to the model through a CPU-GPU direct connection.

Processing workflow: The GPU loads pre-trained AI models (e.g., lightweight CNNs) to perform channel equalization, demodulation, and soft bit generation for IQ data. The inference delay must be controlled within 1000 microseconds to meet Future Network symbol-level real-time requirements. To accommodate floating-point operations in model training, the GPU performs fixed-to-floating-point conversion internally, enabling flexible integration.

Regarding the output interface, the soft bits (LLR) generated by the GPU are transmitted via the PCIe bus to the FPGA in the BBU. The FPGA then forwards the LLR to its internal decoder (e.g., an LDPC decoder) and ultimately feeds the decoding result (CRC) back to the CPU.

The AI receiver model based on BBU (CPU + FPGA + GPU) must adhere to the ONNX format standard, with layered collaboration as its core and testing verification as its safeguard, ensuring the model meets the base station’s real-time, accuracy, and reliability requirements. This process facilitates the transition of AI receivers from the’ lab ‘to the’ live network,’ providing smarter and more reliable reception capabilities for Future Network and future communication networks.

Barrier

The current methods for designing task-specific AI small models in Task 1 and Task 2 suffer from weak generalization capabilities and high design costs.

Key Areas of Focus

Therefore, participants are encouraged to jointly develop a universal large model for both tasks. From a traditional communication perspective, while Task 1’s clutter cancellation and AI receiver differ in overall functional processes, they share overlapping functionalities such as channel estimation. This demonstrates the feasibility of designing and training a universal large model framework for both tasks.

Goal of the Challenge

This Challenge focuses on developing an AI model for Future Network AI PUSCH receivers.

Inference delay: Keep it below 1000 microseconds, with jitter not exceeding ±100 microseconds

Model input data: frequency domain data of the base station’s receiving antenna

Model output data: LLR soft bit data after receiver demodulation

Model features: includes channel estimation, equalization, and demodulation modules of a traditional receiver, but excludes decoding

Supports maximum configuration: 16 antennas, 400MHz bandwidth, 120KHz subcarrier spacing

Performance gain requirement: Compared to traditional receivers, especially those employing LMMSE channel estimation and LMMSE equalization, achieving the same block error rate (BLER) level requires a signal-to-noise ratio (SNR) gain of over 2 dB.

Lightweight requirement: fewer than 500,000 model parameters.

Specific Devices and Capabilities for Task 3

1) Open datasets: Task 1 and Task 2 provide open access to datasets.

2) Open Device Capabilities: The open hardware platform from Task 1 and the open AI model inference base from Task 2 serve as platforms for model deployment and inference experiments.

For each of the 3 Tasks the Future Network Open Test Lab provides applicants with the following enabling devices and capabilities:

• Data
• Functionality
• Interfaces
• Scenarios
• Services

It is implemented through two modules—the Future Network prototype system and the open test environment—with specific content integration achieved via an open collaboration platform.

The facility provides design verification services for fields such as communication, sensing, AI, cloud-based service-oriented solutions, and emerging services, supporting technical roadmap forecasting and design scheme evaluation.

The specific goals for the Challenge are as follows:

1. A dedicated model for Task 1 clutter elimination and Task 2 AI receiver design.
2. A general model for Task 3

The Challenge Process is as follows:

1. Potential Applicants register to the Challenge and remotely receive the data via cloud. Applicant runs / builds their AI model themselves on the data received
2. Applicant then receives a new data set of test verification data for their AI model to run on
3. Applicant sends back to GTI performance results of their AI model and a “demo” of it
4. Challenge judges assess the Round 1 applications and for those which are shortlisted to go through to the final round
5. Applicant is invited to send their AI model to GTI (Beijing) for final round evaluation 

Note: Applicants may choose to participate in one, two, or all three tasks.

Important Dates

Submission of Results: July 10, 2026

Result Coordination: August 15, 2026 (GTI acceptance & evaluation, initiation of device connection)

Completion Date: October 30, 2026 (finalisation of device connection)

Frequently Asked Questions (FAQs)

• Q: Can the lab and the research be located anywhere in the world – or does the lab need to be located in China?
  • A: From GTI’s perspective Applicants can be located anywhere in the world. Applicants should be aware of any local applicable regulations in their own jurisdiction.
• Q: After completion of the Challenge who will own (have the intellectual property rights to) the output/ deliverables?
  • A:  The intellectual property rights to the outputs will be owned by the Applicant