Physical AI Training
A physical training environment for AI coding agents.
Unlike simulated environments, Fieldwork runs on real chips and peripheral devices — exposing agents to the same challenges faced by embedded engineers.
Why Fieldwork
Coding agents underperform on firmware. We're fixing the data.
01
Real Hardware
Timing constraints, communication protocols, hardware variability, noisy signals, and device-specific behavior — all on physical chips and peripherals.
02
Industry Benchmark
Developed in consultation with firmware engineers across robotics, medical devices, aerospace, IoT, manufacturing, and more.
03
Training Data
Firmware tasks are under-represented in training data. Fieldwork fills that gap with structured, verified tasks across the embedded landscape.
Follow Along
Learn about the specific firmware cases we've developed for the benchmark.
We write about embedded engineering challenges, how current models handle them, and what we're learning as we build the benchmark.
Read our blog