Speaker
Description
Recent work has connected large language models to GNU Radio through the Model Context Protocol, enabling natural-language flowgraph generation. These systems can build and validate flowgraphs, but they reason in open loop — the agent never observes whether its design actually works on the air — and prior efforts have found pure LLM orchestration to be sample-inefficient. We present an agentic framework that closes this loop through real radios. An LLM agent constructs transmitter and receiver flowgraphs from GNU Radio's signal-processing primitives via MCP, deploys them to software-defined radios (USRP and ADALM-Pluto), and runs automated hardware-in-the-loop tests over a controlled RF path. Measured physical-layer metrics — bit error rate, EVM, SNR, and constellation and spectrum imagery interpreted by a multimodal model — form the reward signal. The agent then reasons over these observations to iteratively reconfigure modulation, coding, filtering, and synchronization parameters, with explicit memory of which edits improved which metrics. Rather than naively driving BER to zero, the agent optimizes a constrained objective: meeting a BER target while maximizing spectral efficiency under a fixed link budget. A hybrid policy delegates continuous parameter tuning to a classical optimizer while the LLM handles structural decisions, mitigating the cost of slow hardware iterations. We report convergence behavior, iteration counts, and final link performance across modulation scenarios on physical hardware, and discuss reality-grounded feedback as a path toward autonomous, self-optimizing radios. Code and flowgraphs will be released open-source.
| Talk Length | 15 Minutes |
|---|---|
| Acknowledge | Acknowledge In-Person |