A self-driving physics lab built from an Arduino, LEDs and a light sensor

A closed-loop "self-driving" lab does not have to be expensive. A team built one for physics teaching from an Arduino microcontroller, eight LEDs and a light sensor, using only commercially available parts and open-source software. Its job: reproduce a target colour spectrum. Three machine-learning methods compete to do it, so students see their strengths and limits first-hand.

The setup

Figure 1 of the paper, arXiv:2604.13139 (2026), CC BY 4.0; changed here: shown as a picture on a page.

A computer runs the optimization and picks candidate voltages. The Arduino passes them to a digital-to-analog converter, which drives eight LEDs of different wavelengths. A multichannel light sensor (AS7341) reads the result in ten channels — eight narrow visible bands from about 415 nm to 680 nm, plus a clear and a near-infrared channel. The computer compares the measured spectrum with the target and adjusts the voltages, and the loop repeats. Firmware is written in the Arduino IDE and the algorithms in Python with NumPy and PyTorch. The hardware list covers the sensor, converter, Arduino Uno, 3D-printed holders, LEDs, wires, breadboard and cable.

Figure 2 of the paper: the system workflow. arXiv:2604.13139, CC BY 4.0; changed here: shown as a picture on a page.

Three ways to find the right voltages

1. Traversal. A brute-force search through a tree of candidate voltages. A coarse 0.1 V step is quick but only partly matches the target; a finer 0.05 V step matches better and takes longer.
2. Bayesian optimization. A probabilistic model of the voltage-to-spectrum relation, with confidence intervals, sharpens as data arrive. After 400 iterations predicted and measured spectra were nearly indistinguishable.
3. Deep learning. A convolutional neural network trained on a synthetic dataset of about 100,000 voltage–spectrum pairs, made by summing the measured response of each LED at random voltages. It predicts the voltages directly from a target spectrum. Two sizes were tested, with filters 32–64–128 and 64–128–256; the larger matched the target more closely.

Figure 3 of the paper: traversal. arXiv:2604.13139, CC BY 4.0; changed here: shown as a picture on a page.

Figure 4 of the paper: Bayesian optimization. arXiv:2604.13139, CC BY 4.0; changed here: shown as a picture on a page.

Figure 5 of the paper: building the training data. arXiv:2604.13139, CC BY 4.0; changed here: shown as a picture on a page.

Figure 6 of the paper: the neural-network controller. arXiv:2604.13139, CC BY 4.0; changed here: shown as a picture on a page.

How they compare

Time per spectrum, in seconds, as listed in the paper (for the neural network, the first figure is training time):

The paper's takeaways:

• Traversal is simple and cheap to compute, but slow to converge and sensitive to its starting point and to noise. It is a useful baseline.
• Bayesian handles noise well and quantifies uncertainty, but needs more complex computation and careful calibration of priors.
• Deep learning captures nonlinear relationships best and infers quickly once trained, but needs training data and computing power.

The authors suggest choosing by need: traversal for quick deployment with few resources, Bayesian where uncertainty matters, deep learning for high accuracy with plenty of data.

Limits

The sensor's resolution and dynamic range restrict accuracy compared with high-end spectrometers, and LED drift and stray light can affect reproducibility. The authors propose better detectors, active calibration and hybrid controllers. A lab guide, code, firmware and datasets are published openly on Zenodo and GitHub.

Sources

• Yang Liu, Qianjie Lei, Xiaolong He et al., Building an Affordable Self-Driving Lab: Practical Machine Learning Experiments for Physics Education Using Internet-of-Things, arXiv:2604.13139 (2026), licence CC BY 4.0. Changed here: rewritten and condensed in hubnx's own words; the paper's figures are shown as pictures on a page; footnotes and figures under other terms are left out.