Context
FreeFall Aerospace operates a novel spherical-reflector allsky antenna in Tucson, Arizona — capable of tracking any object in the visible sky simultaneously, without mechanical steering. CatSat is a 6U CubeSat from the University of Arizona, launched into LEO with an experimental inflatable antenna and a software-defined radio downlink.
DeepBro deployed DeepLink-02 at the FreeFall site to validate the Cognitive Radio Engine against a real, low-SNR, fast-moving target.
What we did
- Connected DeepLink-02 to the FreeFall allsky feed via direct USB / SoapyRemote
- Scheduled passes against the CatSat TLE using Skyfield
- Captured raw IQ at 2.048 MS/s during each pass window
- Ran the AMR engine inline (sub-50ms inference, Jetson Orin Nano)
- Cross-checked classifications against ground-truth where available
Results
The AMR engine reliably classified the downlink modulation across every viable pass, with settled-state confidence above 96% on QPSK and 16QAM signals at SNRs as low as 8 dB. The full IQ → classification → SNR readout pipeline ran end-to-end without dropped frames.
“The DeepBro engine slotted in beside our existing SDR pipeline without any RF chain changes. The latency made it usable as a live decision input, not just a post-hoc analysis tool.”
What’s next
We’re extending this validation toward adaptive modulation selection — feeding AMR classifications back into the demodulator’s decision logic in real time, and measuring the resulting improvement in usable data per pass.