Research digest, week of 9 November 2026
Eight arXiv papers on indoor multi-band channel measurements, an over-the-air neural 5G receiver, a PlutoSDR radar target simulator, UAV channel data, satellite D2D interference, Wi-Fi 8, LDPC and OTFS.
Measured data was the common thread this week. Three papers took radios out of simulation. A TU Wien group compared indoor MIMO channels at 2.55 GHz, 12.75 GHz and 25.5 GHz with one setup and found the angular structure largely shared across bands, which is useful for anyone extending sub-6 GHz models into the upper mid-band. A neural 5G NR uplink receiver trained on a ray-traced digital twin was tested over the air on USRP B210 radios. It beat a conventional receiver at every position, though only in the building it was trained for. A UAV testbed built on OpenAirInterface showed hardware effects dominating raw channel data from commodity SDRs.
Low-cost test equipment also featured. An unmodified ADALM-Pluto now works as a four-target radar simulator for sensing experiments. On the regulatory side, a cross-border interference study for satellite direct-to-device service in 698 to 960 MHz found coexistence infeasible without mitigation, a timely input for WRC-27. Rounding out the week were Wi-Fi 8 multi-AP scheduling, post-layout results for a 5G LDPC decoder, and a real-time GPU receiver for Zak-OTFS.
Multi-Band Indoor MIMO Channel Measurements Across Sub-6 GHz, Mid-Band and mmWave: Comparing 2.55 GHz, 12.75 GHz and 25.5 GHz #
The authors measured the same indoor lecture room at 2.55 GHz, 12.75 GHz and 25.5 GHz with one setup, so the three bands can be compared directly. A Rohde & Schwarz ZVA24 vector network analyser measured 400 MHz of bandwidth, and single monopole antennas moved by linear positioners formed virtual 4 × 4 arrays at each end, giving a 16 × 16 MIMO channel, with an external mixer chain for the 25.5 GHz band. They extracted 20 specular multipath components per channel with the SAGE algorithm in line-of-sight and obstructed line-of-sight positions. The angular power distributions were similar across all three bands, with the weaker paths changing most with frequency. This is a field measurement in one static room, with the dynamic range clipped to 30 dB for all bands, so it suggests, rather than proves, that sub-6 GHz angular structure is a reasonable starting point for upper mid-band models.
Design and Over-the-Air Evaluation of a Digital Twin-Driven Site-Specific Neural Receiver for 5G NR on NVIDIA DGX Spark #
The authors scanned an indoor corridor, built a digital twin, and used Sionna RT to generate 28,671 ray-traced channels to train a neural receiver for the 5G NR uplink. The receiver replaces channel estimation, equalisation and demapping, and passes soft bits to the standard LDPC transport-block decoder. In a lab testbed with NI USRP B210 radios and NVIDIA DGX Spark computers at 3.30948 GHz, 30 kHz subcarrier spacing, 51 resource blocks and 16-QAM, the site-trained receiver had the highest mean uplink rate at all five test positions. At one non-line-of-sight position it reached 16.2 Mbit/s against 4.1 Mbit/s for a conventional LS/LMMSE receiver and 2.4 Mbit/s for a neural receiver trained on 3GPP channel models. The tests are single-antenna, in one building, and the receiver is by design tied to that site.
A Four-Branch Radar Target Simulator with Fractional Delay on a PlutoSDR #
This paper turns an unmodified ADALM-Pluto into a radar target simulator for sub-6 GHz sensing experiments. FPGA logic generates four independent echoes, each with its own delay, Doppler shift and gain, and fractional-delay filters set range below one sample. Delay and carrier phase are updated together, so a moving target shows the correct frequency-dependent Doppler across an OFDM signal. Over-the-air lab tests with an ESPARGOS radar array reproduced 10 mm range steps with largest median errors of 2.6 mm in overclocked mode and 1.6 mm in standard mode. The fixed hardware latency puts the nearest emulated target at an equivalent 62.1 m (with the interface overclocked to 160 MHz, beyond its 61.44 MHz rating) or 197.7 m at standard settings, and the FPGA source and firmware are on GitHub.
Empowering AI towards 6G: Realistic UAV Channel Data Acquisition using Open-Source Solutions #
The authors built a UAV-to-ground 5G measurement testbed from an OpenAirInterface gNB on a USRP X410 at 3619.2 MHz and a commercial 5G modem on a multicopter. They modified the gNB to log per-subcarrier channel estimates from the uplink sounding reference signal, time-aligned with flight telemetry. About 100 minutes of field flights at 40 to 60 m altitude and up to 15 m/s showed that the dominant spectral ridges in the raw data did not follow the expected kinematic Doppler unless the tracking was guided, which the authors attribute to hardware effects such as phase noise. Deep-learning channel predictors trained on this data behaved differently from what idealised channels suggest, with a Transformer doing best in their case study. The study covers one hardware chain, a single link and low speeds, and the testbed and dataset are released.
Coexistence of Direct-to-Device (D2D) Satellite Systems with 5G/6G Terrestrial Networks: Cross-Border Aggregate Interference Analysis in Saudi Arabia #
This study asks whether satellite direct-to-device downlinks serving Jordan and Iraq in the 698 to 960 MHz band would interfere with terrestrial IMT phones in Saudi Arabia, a question under WRC-27 agenda item 1.13. It is a Monte Carlo simulation using ITU-R antenna and propagation models, two hypothetical constellations at 500 and 525 km altitude, and victim phones spread along the real borders up to 40 km inland. The protection criterion was exceeded up to 88% of the time, needing up to 27 dB of extra isolation, and random satellite selection was worse than highest-elevation selection. A one-cell border isolation zone helped only moderately, while a narrow-beam, deep-null phased-array pattern cut exceedance to about 40%. The systems are hypothetical, so the numbers are inputs to regulatory study rather than predictions for a real constellation.
Machine Learning for Scalable Multi-AP Coordination in IEEE 802.11bn Networks #
This short paper summarises a line of work on coordinated spatial reuse for IEEE 802.11bn (Wi-Fi 8), where the hard part is choosing which access points transmit together, to which stations, at what power and with which MCS. It covers hierarchical multi-armed bandits, mixed-integer programs that give upper bounds, and a generative pipeline using graph neural networks and flow matching that produces a schedule in under one second for more than 30 access points. In the authors' open-source Monte Carlo simulator the bandit approach raised aggregate throughput over legacy DCF by 80% on average in random scenarios, and agents were also run on an openwifi SDR testbed. Most numbers are simulation results drawn from the authors' earlier papers, and the schemes go beyond the two-AP coordination that 802.11bn currently allows.
Optimization of Finite Alphabet Message Passing Decoding on Fully Node-Parallel Architectures for 5G LDPC Codes #
Finite-alphabet LDPC decoders replace arithmetic in the message updates with small lookup tables, which saves wires and power in fully parallel hardware. The authors design these tables with Bayesian optimisation against frame error rate, without assuming a channel or modulation, and produce one set of tables used in every iteration. For the (1000, 500) base graph 2 5G NR LDPC code on a simulated binary-input AWGN channel, the decoder matched a normalised min-sum decoder in error rate. After synthesis and place and route in a 12 nm FinFET process, it reached 21.6 Gbit/s at 15.36 pJ/bit against 25.02 pJ/bit for min-sum, a 38.6% gain in energy efficiency. These are post-layout estimates for one code, not measured silicon.
Real-Time Low-Complexity Zak-OTFS with Superimposed Spread Pilots on GPUs #
Zak-OTFS with spread pilots overlays the pilot on the data across the whole delay-Doppler grid, which saves resources but makes channel estimation and pilot cancellation expensive. Using the chirp structure of the pilot, the authors reduce the cross-ambiguity calculation to one 2D IFFT plus element-wise operations, over 7000 times faster than a partly vectorised direct GPU implementation at a frame size of 2053 by 31. Their full receiver pipeline with one turbo iteration fits a 1.03 ms latency budget on an NVIDIA RTX 6000 Ada GPU. Larger frames give more processing gain, so pilot power can drop and the bit error rate is close to converged after one iteration. All results use a simulated ITU Vehicular-A channel, with no radio in the loop.