In unsourced random access, many cheap wireless sensors send short messages at the same time using one shared codebook. The base station must recover the unordered list of messages from their overlapping sum, without knowing who transmitted. Neural receivers for this problem are fast. The masked-diffusion decoder CIDER, for example, runs more than 100× faster than classical belief propagation at blocklength 48. But these receivers have only been trained and tested on idealized synthetic channels. Real radios add impairments that such simulators leave out: frequency offsets, phase noise, amplifier nonlinearity, and unsynchronized arrival times.
Robotics faced the same gap between simulator and hardware and addressed it with domain randomization, which means training across a deliberately wide range of simulator settings so the learned model stops depending on any single one. This note proposes bringing that approach to multi-user radio decoding with one addition that only this setting offers. The error-correcting code gives a free, label-free signal of how well the receiver works on real signals, because decoded messages either satisfy their parity checks or they do not. We propose to use that signal to reshape the randomization distribution itself, rather than to fine-tune the network on real data. The receiver never sees a labeled real frame. Yet its training distribution moves toward the impairments actually present, which avoids the over-conservative behavior that very wide randomization tends to cause.
Core Claim
The claim has two parts. First, a neural receiver for unsourced random access can be trained entirely in a randomized channel simulator and still decode real, impaired signals without further training. Second, and more important, the useful way to use unlabeled real traffic is to infer simulator parameters from code-structured statistics of the receiver's own outputs, not to self-train the receiver. These statistics include syndrome weights of decoded rows, the parity-pass rate at each confidence level, and the residual energy left after cancelling decoded users.
We predict that this "parity-guided" randomization will do better than both uniform wide randomization and pseudo-label fine-tuning. The gain should be concentrated on the weakest users. The reason is that pseudo-labels exist only for users who were already decoded correctly. Fine-tuning on them therefore reinforces the easy regime. A re-fitted simulator, by contrast, can produce labeled examples in the hard regime, where no real labels ever appear.
What Exists and What Is Missing
Three papers supply the receiver side. Structured Masked Diffusion for Joint Multiuser Decoding (CIDER) decodes many overlapping users by filling in a grid of masked symbols. It uses a module that makes hypothesis rows compete for signal energy and a module that injects the LDPC parity graph. Its authors state that they assume unit-gain, unfaded channels with no radio impairments. We reuse its architecture and remove exactly that assumption.
A Polar Code Based TIN-SIC Scheme for the Unsourced Random Access in the Quasi-Static Fading MAC is the strongest classical receiver under fading. It decodes the strongest user while treating the others as noise, estimates channels blindly by clustering, and then subtracts decoded users one at a time. Its authors also acknowledge that errors early in the subtraction chain propagate to later stages. Pilot-Free Unsourced Random Access Via Dictionary Learning and Error-Correcting Codes uses parity checks to resolve the permutation and phase ambiguity of blind factorization. That is the closest precedent for treating the code as a validator. It uses the code inside a single decode, however, not as a signal for adapting a training distribution, and it assumes perfect synchronization and i.i.d. Rayleigh channels.
Two other papers motivate the problem. Unsourced Random Access: A Comprehensive Survey reports that receivers degrade under hardware distortion and asynchrony. Massive Uncoordinated Access With Random User Activity shows that not knowing the number of active users costs an optimal decoder only about 0.5–0.7 dB. That result sets a target a randomized neural receiver should approach when the user count is itself one of the randomized parameters.
On the randomization side, Sim-to-Real: Learning Agile Locomotion For Quadruped Robots and (CAD)$^2$RL show that broad randomization transfers without real data, even when the simulator's fidelity is low. The quadruped paper also found that system identification alone was not enough. Active Domain Randomization and Domain Randomization via Entropy Maximization reshape the randomization distribution using signals computed in simulation. Neural Posterior Domain Randomization fits a posterior over simulator parameters from real observations. That is the closest method to ours. It assumes the real observations are directly informative about the parameters. Here the raw superposed waveform is a poor summary statistic, and the proposal's specific bet is that code-structured statistics carry the information. No paper in the file joins these two lines.
Research Program
We would first build a randomized simulator for the multi-antenna and single-antenna unsourced channel. Each user $k$ contributes a codeword passed through a complex gain drawn from a Rician-to-Rayleigh family. We then add a per-device frequency offset, a phase-noise process, a memoryless amplifier nonlinearity, a fractional timing offset in $[0, T_s)$ where $T_s$ is the symbol period, and receiver quantization. The number of active users is Poisson with a randomized rate. Collect all of these impairment parameters in a vector $\xi$, drawn from a distribution $p_\phi(\xi)$ with parameters $\phi$.
The receiver extends CIDER with a per-row complex gain estimate and a timing-shift embedding, so the competition between rows can operate on unequal amplitudes. It uses non-binary LDPC codes over GF(64), as in CIDER. During training we can supervise an auxiliary head with the true simulated channels, which borrows the privileged-information idea from Asymmetric Actor Critic for Image-Based Robot Learning. That head is dropped at deployment.
The adaptation loop proceeds as follows. We train the receiver under a broad $p_{\phi_0}$ and run it on unlabeled real frames $\{y_j\}$. From its output we compute a statistic vector $s(y)$: the histogram of syndrome weights across hypothesis rows, the parity-pass rate within each confidence decile, the number of distinct parity-valid messages, and the spectrum of residual energy after cancelling valid messages. The same statistics are computed on simulated frames drawn at candidate values of $\xi$. A neural posterior estimator $q$ is trained on these simulated pairs, and the randomization distribution is updated toward the pooled posterior:
$$ p_{\phi_{t+1}}(\xi) \propto \frac{1}{J}\sum_{j=1}^{J} q\big(\xi \mid s(y_j)\big), $$
Here $J$ is the number of unlabeled real frames. We then broaden this distribution by a fixed entropy floor so the receiver stays robust to device-to-device spread, and we continue training the receiver in simulation under $p_{\phi_{t+1}}$.
One design choice needs a reason. Syndrome weights and parity-pass rates respond to impairments in a code-specific way. Frequency offset rotates symbols and breaks parity in a pattern different from amplifier compression, for instance. Raw moments of the received signal mostly measure total power. The statistics should also resist being gamed: a random row passes $r$ parity symbols over GF(64) with probability about $64^{-r}$, so counts of valid messages are nearly free of false alarms.
Baselines
The naive reference point is CIDER trained on its original setting, a fixed channel without impairments, and evaluated on impaired signals. It shows how large the gap between simulator and reality is. A second reference is a receiver trained on a single "nominal realistic" channel, meaning Rayleigh fading plus the impairment values from calibration measurements. This is the system-identification-only approach, and it is what the quadruped paper found insufficient in robotics.
The strongest classical competitor is the Polar TIN-SIC receiver of A Polar Code Based TIN-SIC Scheme. In quasi-static Rayleigh fading with blocklength 512 it reports resolving "up to 14 simultaneous users per slot with a per-user error probability below 10%." It is the strongest because it was designed for fading without channel knowledge and does not depend on a learned training distribution. We will run it with the authors' configuration and add standard frequency-offset and timing estimation front ends, so it does not lose simply for lacking obvious engineering. Beating it means supporting more than 14 users per slot at per-user error below 10% on impaired channels, where we expect its cancellation chain to degrade, while staying within about 1 dB of it on the clean Rayleigh channel. The comparison is only fair at matched spectral efficiency, so the neural receiver's code and slot structure will be set to the same payload size and number of channel uses.
The strongest neural competitor is CIDER with uniform wide randomization over the same parameter ranges. It gets the same architecture, training steps, and hyperparameter search budget as our method. Beating it means lower required energy per bit at per-user error 0.05, and above all lower error on the weakest quartile of users, with no loss on the nominal channel.
Two further neural baselines complete the comparison. Entropy-maximizing randomization in the spirit of Domain Randomization via Entropy Maximization is adapted using simulated signals only. Self-training fine-tunes the uniform-randomization receiver on real frames, using parity-valid decoded messages as pseudo-labels.
Two ablations isolate the key ingredient. The first replaces the code-structured statistics with generic received-signal statistics: covariance eigenvalues, amplitude histograms, and spectral moments. This is essentially Neural Posterior Domain Randomization applied to radio. The second is an oracle that trains under the true held-out impairment distribution and serves as an upper bound. If the generic-statistics variant matches ours, the code-structured signal is not what matters, and the claim fails.
Why not just…?
Why not just fine-tune the receiver on parity-valid real decodes? This is the most serious alternative. The research file's claim that real labels do not exist is wrong, because parity-valid messages are near-certain labels. The problem is where the labels come from. They come from users who were already decoded, and those users are disproportionately strong and lightly impaired. Weak users, under the worst frequency offset or amplifier compression, never produce labels. We predict self-training improves the average error while leaving the weakest-quartile error flat. The self-training baseline, broken down by user strength and impairment severity, tests this directly. If self-training matches us on weak users, we will report parity-guided randomization as an initialization for fine-tuning rather than a replacement for it.
Why not just calibrate the radios and simulate them faithfully? In a lab this is easy: send known frames and fit the impairments. Deployment is different. Each cheap device has its own frequency offset and amplifier, and these drift. The survey stresses that per-device calibration across very large populations is impractical. The quadruped work also found system identification alone insufficient even for a single robot. Our test calibrates one subset of transmitters, then evaluates on transmitters that were not calibrated, compared against our method, which fits the population distribution from unlabeled traffic from all devices.
Why not just use a classical receiver with explicit impairment estimation? Model-based receivers with frequency-offset and timing estimation are a real threat on mild impairments, and TIN-SIC may win there. The case for the neural route rests on two things. CIDER reports being 6×–100× faster than FFT-based belief propagation. Successive cancellation, meanwhile, is fragile to residual estimation error, which the TIN-SIC paper itself acknowledges. We sweep impairment severity to locate the crossover rather than assume there is one.
Why not just randomize as widely as possible? Wide randomization is known to produce conservative policies. In decoding, the analogue is a receiver that spends capacity on impairments that never occur and loses energy efficiency on the channels that do. The uniform-randomization baseline, measured on the nominal channel, quantifies that cost.
Proposed Experiments
The first tier is controlled and entirely in software. "Reality" is a held-out simulator whose impairment models lie outside the training family. It uses a Wiener phase-noise process where training used i.i.d. jitter, a Rapp-type amplifier curve where training used a cubic model, and clustered arrival times. Because the true parameters are known, we can check whether the posterior concentrates near them.
The second tier is over the air. Several software-defined radios act as sensors and transmit to a multi-antenna receiver radio, with ground truth logged on the transmit side for evaluation only. The file lists no over-the-air benchmark for this problem, so this testbed has to be built. It is the main engineering cost.
Evaluation follows the unsourced-access literature. We report per-user error probability against energy per bit, the maximum number of users per slot at per-user error ≤ 0.10, the energy per bit needed for per-user error 0.05, and false-alarm rate. We add weakest-quartile per-user error, and the gap to the known-user-count case, compared against the 0.5–0.7 dB reported for optimal random coding. Test conditions range from AWGN to single-antenna quasi-static Rayleigh fading to 64-antenna fading, the configurations used across the cited papers. Neither the file nor CIDER gives a numerical symbol or codeword error rate, so the neural comparison will rest on our own re-runs. For CIDER, the file lists no reported error-rate figure on Lacuna.
Three decisive checks settle the claim. The first is whether parity-guided randomization beats uniform randomization on weakest-quartile error at matched compute. The second is whether its advantage disappears when code-structured statistics are replaced with generic ones. The third is whether it beats self-training on the users that self-training never labels. If all three hold in both tiers, the code has served as its own sim-to-real sensor. If only the first holds, the contribution reduces to randomized training of neural receivers for this problem, which is still useful but a much smaller result.
References
- Structured Masked Diffusion for Joint Multiuser Decoding
- A Polar Code Based TIN-SIC Scheme for the Unsourced Random Access in the Quasi-Static Fading MAC
- Pilot-Free Unsourced Random Access Via Dictionary Learning and Error-Correcting Codes
- Unsourced Random Access: A Comprehensive Survey
- Massive Uncoordinated Access With Random User Activity
- Sim-to-Real: Learning Agile Locomotion For Quadruped Robots
- CAD2RL: Real Single-Image Flight Without a Single Real Image
- Active Domain Randomization
- Domain Randomization via Entropy Maximization
- Neural Posterior Domain Randomization
- Asymmetric Actor Critic for Image-Based Robot Learning