⚡ electrical engineering

A Transformer for Joint Multi-Receiver Pilotless Wi-Fi Decoding

This paper proposes a fully pilotless multi-access-point Wi-Fi receiver utilizing a self-attention Transformer that, by leveraging spatial diversity across multiple access points, achieves practical bit error rates and higher spectral efficiency comparable to ideal channel-aware baselines.

Original authors: Xavier Tardy, Grégoire Lefebvre, Apostolos Kountouris, Haïfa Farès, Amor Nafkha

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xavier Tardy, Grégoire Lefebvre, Apostolos Kountouris, Haïfa Farès, Amor Nafkha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Wireless signals travel through the air like invisible ripples, carrying our messages from phones to routers. To ensure these messages arrive intact, engineers have long relied on a system of built-in checkpoints called "pilots." These are small, known signals sent alongside the actual data, allowing the receiver to measure how the air distorted the message and correct it. This method works well, but it comes with a cost: every bit of space used for a pilot is a bit of space taken away from the actual message. In the crowded, complex environments of modern offices and homes, where many devices compete for the same airwaves, this overhead becomes a significant bottleneck. The question facing researchers is whether it is possible to strip away these pilots entirely and still decode the message perfectly, relying instead on the receiver's ability to learn the shape of the signal from the data itself.

A team of researchers from France has taken a significant step toward answering this question by developing a new type of Wi-Fi receiver that operates without any pilots at all. Their work, presented at a major international conference in Krakow, focuses on a scenario where multiple access points—essentially the routers or base stations—listen to the same transmission from a single device. Instead of trying to measure the channel with dedicated pilot signals, their system uses a sophisticated artificial intelligence model, specifically a type of neural network known as a Transformer, to analyze the raw, complex patterns of the incoming signals. This model treats every tiny piece of data from every listening station as a unique clue, weaving them together to reconstruct the original message. The researchers found that while a single listening station struggles to make sense of the signal without pilots, a cooperative network of just three or more stations can decode the message with remarkable accuracy, even outperforming traditional systems that rely on perfect knowledge of the channel.

The core of this innovation lies in how the system processes information. In a standard Wi-Fi setup, the receiver first estimates the condition of the wireless path using pilots, then uses that estimate to clean up the data. This new approach skips the estimation step entirely. The researchers designed a system that takes the raw, noisy electrical signals received by multiple antennas and feeds them directly into a neural network. This network is built to recognize patterns across both time and space. It looks at the signal received by each antenna on each frequency channel simultaneously, learning how the different antennas see the same message from different angles. By combining these diverse perspectives, the system can infer the distortions caused by walls, glass partitions, and other obstacles without ever needing a pre-sent pilot signal to tell it what to expect.

To test this idea, the researchers did not rely on simple mathematical models but instead simulated a realistic indoor environment using advanced ray-tracing software. They created a virtual office space with walls and glass screens, placing a single user in the center and up to five access points around the room. In this simulation, the signals bounced off surfaces and diffracted around corners, creating a complex web of interference that mimics real-world conditions. The team trained their neural network on millions of these simulated scenarios, teaching it to translate the raw, distorted signals into the correct sequence of ones and zeros. Crucially, the network was trained without any knowledge of the specific channel conditions or the presence of pilots, forcing it to learn the underlying structure of the data purely from the patterns it observed.

The results of these simulations revealed a clear threshold for success. When the system used only a single access point, it failed to decode the message reliably, with error rates remaining high even when the signal was strong. This confirmed that without pilots, a single viewpoint is simply not enough to untangle the complex distortions of an indoor environment. However, as soon as the system was given a second viewpoint, the performance improved dramatically. With three access points working together, the system achieved error rates low enough for practical use, reaching a level of reliability where only one in a hundred thousand bits was incorrect. This performance was achieved while using the entire bandwidth for data, effectively boosting the speed of the connection by eliminating the overhead of pilot signals.

Perhaps the most striking finding was that this pilotless system, when equipped with three or more access points, performed better than a traditional system that had perfect, ideal knowledge of the channel conditions. In the simulation, the traditional system was given the luxury of knowing exactly how the signal was distorted, yet it could not match the accuracy of the neural network that learned to navigate the distortion on its own. This suggests that the neural network's ability to fuse information from multiple sources is so powerful that it can overcome the lack of explicit channel measurements. The system proved robust across many different random starting conditions, indicating that the solution is not a fluke but a stable, learnable capability.

This work demonstrates that the future of high-speed wireless communication may not require more complex signaling or more pilots, but rather smarter receivers that can collaborate. By allowing multiple access points to share their observations and process them through a unified neural network, it becomes possible to strip away the safety net of pilot signals and reclaim that bandwidth for data. While the current results are based on simulations in a controlled virtual environment, they provide a strong proof of concept that fully pilotless, multi-antenna Wi-Fi is not just a theoretical possibility but a practical reality waiting to be deployed. The path forward involves refining these models to handle multiple users simultaneously and simplifying the architecture for real-world hardware, but the fundamental barrier of needing pilots for reliable decoding has been shown to be surmountable.

Technical Summary: A Transformer for Joint Multi-Receiver Pilotless Wi-Fi Decoding

Problem Statement
Future Wi-Fi generations (e.g., IEEE 802.11be/bn) aim to maximize spectral efficiency in dense deployments through multi-link coordination. While reducing pilot overhead is attractive—particularly for uplink scenarios where multiple Access Points (APs) observe the same transmission—removing pilots entirely introduces significant decoding challenges. In frequency-selective indoor channels, the absence of explicit Channel State Information (CSI) leaves receivers vulnerable to amplitude/phase distortions and deep fades.

Conventional approaches face two main limitations in this context:

Pilot-based pipelines: These estimate per-AP channels and combine soft metrics (e.g., Maximum Ratio Combining). However, they fail to fully exploit frequency-selective reliability differences across APs and suffer from spectral overhead.
End-to-end learned schemes: Previous neural approaches often require joint training of the transmitter and receiver (e.g., learned constellations), which is incompatible with standardized Wi-Fi waveforms where the transmitter is not under the receiver's control.

The core challenge addressed is achieving reliable, fully pilotless decoding for a standardized Wi-Fi uplink using only receiver-side processing, without explicit CSI or dedicated pilots.

Methodology
The authors propose a fully receiver-side neural architecture based on a self-attention Transformer that processes raw complex frequency-domain observations from multiple APs. The system model assumes a single-antenna User Equipment (UE) transmitting a coded OFDM symbol to $N_R$ single-antenna APs.

Input Representation: The receiver constructs tokens for every (AP, subcarrier) pair. Each token $u^{(r)}_f$ is a 3-dimensional real-valued vector containing the real part, imaginary part, and noise variance estimate of the received sample.
Architecture:
- Embedding: Tokens are linearly embedded into a model dimension ( $d_{model}=64$ ) and augmented with fixed sinusoidal positional encodings based on the subcarrier index to preserve frequency structure.
- Encoder: A stack of 4 Transformer encoder layers processes the tokens. Each layer utilizes multi-head self-attention over the full set of tokens, enabling the model to capture both wideband correlations across subcarriers and spatial correlations across APs.
- Fusion & Output: The encoded representations are fused per subcarrier by averaging across APs (a permutation-invariant operation). A Multi-Layer Perceptron (MLP) then maps this fused representation to bit-wise logits (Log-Likelihood Ratios, LLRs) for the $M$ bits of the QAM symbol.
Training & Decoding: The Transformer is trained by minimizing bit-wise binary cross-entropy (BCE) against uncoded bit labels. Crucially, the LDPC decoder is not part of the training loop; the network outputs soft LLRs compatible with a standard rate-2/3 LDPC decoder, which is applied only at test time. This ensures the receiver does not adapt to a specific code but learns a general channel representation.

Key Contributions

Pilotless Multi-AP Decoding: The paper demonstrates a purely receiver-side neural design that eliminates the need for dedicated pilots while maintaining compatibility with standardized QAM waveforms and LDPC codes.
Content-Aware Fusion: By utilizing self-attention, the model learns to fuse information across APs and subcarriers without explicit channel estimation, effectively handling frequency-selective fading and deep fades through spatial diversity.
Spectral Efficiency: The approach removes pilot overhead entirely, directly enhancing the net data rate compared to pilot-based baselines.

Experimental Results
The system was evaluated in a realistic ray-traced indoor environment ("Box–Two–Screens") using Sionna RT, with 16-QAM modulation and $N_c = 72$ subcarriers. The number of cooperating APs ( $N_R$ ) varied from 1 to 5.

Single-AP ( $N_R=1$ ): The Transformer-based receiver failed to achieve practical error rates, with a coded Bit Error Rate (BER) remaining around $10^{-1}$ even at high SNR ( $E_b/N_0 \approx 10$ dB). This confirms that inferring a reliable channel from a single noisy observation without pilots is infeasible.
Cooperative Configurations ( $N_R \ge 2$ ):
- With $N_R=2$ , the system entered a "waterfall" region, achieving a BER of $10^{-3}$ at $E_b/N_0 \approx 8$ dB, comparable to a single-link benchmark with perfect CSI.
- With $N_R=3$ , the system reached the same BER ( $10^{-3}$ ) at roughly $E_b/N_0 \approx 6\text{--}7$ dB, corresponding to a gain of a few dB compared to the perfect-CSI single link.
- With $N_R=4$ , the system achieved a coded BER of approximately $10^{-5}$ around $E_b/N_0 \approx 6$ dB.
- With $N_R=5$ , similar reliability ( $10^{-5}$ ) was achieved at lower SNR ( $E_b/N_0 \approx 4$ dB), and the BER further decreased to about $10^{-8}$ near $E_b/N_0 \approx 6$ dB.
- Notably, for $N_R \ge 2$ , the pilotless multi-AP receiver consistently outperformed the idealized single-link perfect-CSI benchmark across the tested range.
Robustness: Performance curves showed narrow standard deviations across independent training runs, indicating robustness to initialization and training noise.

Significance and Claims
The paper claims that fully pilotless decoding is feasible in realistic, frequency-selective Wi-Fi scenarios provided that spatial diversity is available (i.e., multiple cooperating APs). The study establishes that a purely receiver-side Transformer can replace traditional channel estimation and combining pipelines, achieving error rates comparable to or better than ideal benchmarks while eliminating pilot overhead.

The authors position this work as a "first step" toward pilotless multi-user uplinks. They explicitly note that future research is required to validate these results in multi-user settings and to explore simpler attention mechanisms (e.g., axial or structured attention) to reduce computational complexity for potential online, embedded deployments. The work does not claim immediate deployment readiness but rather validates the theoretical and practical viability of the approach in a controlled, realistic simulation environment.

More like this