Train it on your computer.
Prove it on the chip.
Most tools stop at "export code". VeriAI Studio uploads that code to a real Arduino or ESP32 board, sends it inputs over USB, and checks every answer against your model, with the time per prediction measured on the board. With Controller in the loop, your controller runs on the board and controls a plant simulated in VeriAI Studio.
From the lab to the processor, step by step.
Each loop answers one question: does the model work, does the exported code compute the same thing, does the real board do it in time, and does your controller hold the plant when it runs on the board? Click a loop to see how it works in VeriAI Studio.
-
MIL
Model in the loop
Train and evaluate the model in VeriAI Studio: live loss curves, accuracy and confusion matrix on your own data.
The model is trained and checked inside VeriAI Studio, on your own data, before any code is written for a board.
-
SIL
Software in the loop
Before any upload, the exported code's 32-bit and 8-bit arithmetic runs on your computer and is compared with the model. Size and accuracy are checked; no board needed.
Before anything goes to a board, VeriAI Studio works through the exported code's arithmetic on your computer and compares it with your model, row by row.
-
PIL
Processor in the loop
VeriAI Studio uploads the code to your Arduino or ESP32 board, sends test rows (or values you type) over USB and compares every answer. Time per prediction is measured on the chip.
Your model runs on the real processor of an Arduino or ESP32 board. VeriAI Studio sends the inputs over USB and checks every answer against the computer's.
-
CIL
Controller in the loop
Your controller runs on a real Arduino board and controls a plant simulated in VeriAI Studio, step by step over USB. Its response is compared with the software run.
Your controller runs on a real Arduino board. The plant it controls is simulated in VeriAI Studio, and the two talk over USB at every step: the plant sends the setpoint and its measurements, the board answers with the control signal.
MIL checks the model
- Accuracy, error and confusion matrix on your test data
- Live loss curve while it trains
- Compare architectures and settings
SIL checks the exported code
- The board's 32-bit arithmetic, calculated on your computer
- The 8-bit (int8) version: accuracy kept and memory saved
- Model size in float and in 8-bit, before any upload
PIL checks the real processor
- Same answers as the computer, row by row
- Time per prediction, measured on the chip
- Program and working memory used on the board
CIL checks the controller
- Same control signal as the software run, step by step
- Rise time, overshoot, settling time and IAE with the board in the loop
- Time per step on the chip and, in real time, the USB delay
Type a value. The board answers.
Connect the board once, then type any input values in VeriAI Studio and press Send to board. The board's answer appears next to the computer's, with a ✓ when they agree and the time it took on the chip.
- Try values outside your test set, edge cases or a reading you have in mind
- Send on every change answers as you type
- Fill with a test row shows the true answer next to the board's
- Every input is kept in a list you can download as CSV
Real numbers from an Arduino Nano 33 BLE.
Measured with VeriAI Studio's Test on device on an Arduino Nano 33 BLE (nRF52840, Cortex-M4F at 64 MHz). Your numbers depend on the model and the board.
| Model | Lab | Same as computer | Time on the board |
|---|---|---|---|
| Linear regression | ML Lab | ✓ all rows | 5 µs |
| Neural network, 4 → 16 → 3 (iris) | ML Lab | ✓ all rows | 33 µs |
| Neural network, 6 → 16 → 3 · 8-bit | ML Lab | ✓ all rows | 46 µs |
| Random forest, 20 trees | ML Lab | ✓ all rows | 58 µs |
| Deep feed-forward, 4 → 16 → 16 → 16 → 3 | Deep Learning Lab | ✓ all rows | 157 µs |
| Deep network, 4 hidden layers (tanh, GELU, swish) | Deep Learning Lab | ✓ all rows | 658 µs |
| Same deep network · 8-bit | Deep Learning Lab | ✓ all rows | 693 µs |
Classes are identical on the board and the computer; numeric answers agree within 0.1 % (the board's processor rounds floating-point results slightly differently). Times exclude the USB transfer.
Your controller on the board. The plant in VeriAI Studio.
Pick a plant (DC motor, water tank, heater, buck converter or an inverted pendulum), choose the controller (a PID, your own C code, a Fuzzy Lab system or a model trained in ML Lab or Deep Learning Lab) and put it on the board. At every step the plant sends the setpoint and its measurements over USB and the board answers with the control signal. It is in all 9 simulation tools.
| Controller on the board | From | Plant | Same as the software run | Time per step on the board |
|---|---|---|---|---|
| Fuzzy system | Fuzzy Lab | DC motor | ✓ all 301 steps | 646 µs |
| Trained model | ML Lab | DC motor | ✓ all 301 steps | 13 µs |
| Trained neural network | Deep Learning Lab | DC motor | ✓ all 301 steps | 27 µs |
| PID, tuned | Optimizer Lab | DC motor | ✓ all 301 steps | 12 µs |
| Your own C code | DAQ & WiFi Studio | DC motor | ✓ all 301 steps | 12 µs |
| PID | Controller in the loop | Inverted pendulum on a cart (angle and 3 sensors) | ✓ all 501 steps | 13 µs |
| PID | Controller in the loop | Buck converter | ✓ all 401 steps | 12 µs |
| Your own C code | Controller in the loop | Heater | ✓ all 1501 steps | 12 µs |
Measured on an Arduino Nano 33 BLE on 15 Nov 2026. The board's control signal agreed with the software run within 0.00005 % (the board's processor rounds the last digit differently). In real time, with a step every 10 ms, the USB round trip was about 1.4 ms (4 ms for the pendulum's longer messages), no answer was late and the pendulum stayed balanced.
Everything it takes to trust a model on a board.
10 kinds of ML model
Neural network, linear, polynomial, linear SVR, logistic, naive Bayes, decision tree, random forest, k-means and PCA from ML Lab.
Deep networks of any depth
Deep Learning Lab feed-forward networks with every activation in the lab, exported with their real trained weights and input scaling.
Float or 8-bit
See what an int8 model costs in accuracy and saves in memory before you upload it.
Exact-model check
Each test program carries a fingerprint of the model, so a board holding another model is never tested by mistake.
Speed and memory
Time per prediction measured on the chip, program and working memory from the build, against a speed you set.
Reports
A pass / fail verdict, a report file and a CSV with every input, both answers and the time.
Good to know.
Is this hardware-in-the-loop (HIL)?
Test on device is processor-in-the-loop (PIL): the model's code runs on the board's real processor, and VeriAI Studio sends the inputs and reads the answers over USB. Controller in the loop goes one step further: the board runs your controller and closes the loop with a plant simulated in VeriAI Studio, step-locked or in real time over USB. Full HIL would also connect the board's pins to real-time input and output hardware.
Which controllers can run in the loop?
A PID (the gains change without uploading again), your own C code, the fuzzy system from Fuzzy Lab, or a model trained in ML Lab or Deep Learning Lab. Optimizer Lab can tune the PID first, and a run can become training data for a learned controller.
Which boards can I use?
Arduino and ESP32 boards supported by arduino-cli, connected by USB. The measurements on this page are from an Arduino Nano 33 BLE.
Do I need extra hardware?
No. The board and its USB cable are enough. VeriAI Studio builds and uploads the test program for you.
Does testing change my model?
No. The test only reads your trained model. When you are done, one button uploads your own model program to the board again.
Which plan includes it?
Test on device is part of ML Lab and Deep Learning Lab, and Controller in the loop is in the 9 simulation tools (Fuzzy, ML, Deep Learning, Optimizer, XAI, Python Notebook, Dimension Reduction, Agentic AI, DAQ & WiFi) of VeriAI Studio Prime, for Windows and macOS.
Put your next model on a real chip.
Train it, check it, and prove it on your own board, all in VeriAI Studio.