Boards & Bots
Menu
Home
Sign inSubmit a link

NewsBB-0054

AutoMCU: a multi-agent system that designs neural networks to fit STM32 microcontrollers

Penglin Dai and co-authors describe AutoMCU, an LLM multi-agent system that designs neural networks to fit a microcontroller's RAM and flash limits. Vendor toolchain checks reject candidates that would not fit before any training. The paper reports competitive CIFAR accuracy after 1 to 2 hours of customization and tests on several STM32 parts; no code is linked.

Open original source arxiv.org · Penglin Dai et al.

Claimed
Why this evidence label?

Abstract reports real-device deployments on multiple STM32 microcontrollers; boards are not named on the abstract page and no code is linked.

How we check entries
Preview image for AutoMCU: a multi-agent system that designs neural networks to fit STM32 microcontrollers
The source's own preview image, from arxiv.org.

Briefing

Requirements go in as natural language together with the target's hardware specs. Separate agents handle proposal, training, evaluation and deployment, and the authors coordinate them with what they call state-isolated scheduling.

The paper compares AutoMCU with ColabNAS and GENIUS on CIFAR-10 and CIFAR-100. It sets its 1 to 2 hours against hundreds of GPU hours for hardware-aware neural architecture search baselines built for MCUs. The six authors submitted it on 20 May 2026. The abstract does not name the STM32 boards used in the real-device tests.

Written by the Boards & Bots curator

Tags

Board

Report a problem with this entry