Professional engineering · SERES · 2026
From physical touch
to learned understanding.
Magnetic tactile sensing
& machine learning
I worked across the experiment and the model: building the test bed, automating measurements, characterizing the sensor, and leading the neural-network work to estimate contact position and force.
- Setting
- Professional tactile-sensing research
SERES · June–August 2026 - My contribution
- Test-bed design and automation
Measurement methodology and ML - The question
- Where is the contact,
and how much force is applied?
Across the system
The model begins
with the experiment.
01 / Experiment
First, make touch repeatable.
A magnetic tactile sensor turns physical contact into a changing signal. To learn from that signal, the contact itself has to be controlled. I designed and built the automated test bed, connecting the sensor, positioning, force control, and measurement into one experiment.
Automating Torbal force control made controlled contact part of the collection process. The test bed was an essential part of the ML work: it determined what the model could learn from.
02 / Measurement
Give each measurement a meaning.
Automation connects the sensor signal to the contact position and applied force. Characterization then asks a more demanding question: what does a change in the signal actually tell us about the contact?
I developed spatial-resolution and output-reference-point methodology, bringing the measurement process and its labels into the same frame. The result was a dataset grounded in a controlled physical experiment.
03 / Learning
One signal. Two physical questions.
Where is the contact? How much force is being applied? I led the neural-network work that maps magnetic measurements to those two outputs: contact position and force.
The two-headed network makes the relationship visible: a shared interpretation of the sensing signal supports two related predictions. Model improvement stayed connected to how the data was collected, characterized, and evaluated.
04 / Evaluation
Test what the model has learned.
Training fit was one part of the story. Evaluation used two training runs and a separate holdout run, with millions of data points at each stage. MAE, RMSE, and model fit were used to assess prediction quality within the research setup.
The reported improvement in full-scale error was from approximately 50–70% to below 1%, with reported R² around 0.99. These are results from the professional research work, separate from the illustrative browser study below.
Reported research results
Better measurements.
Better predictions.
Results within the research setup. The browser study uses its own synthetic model and data.
- Full-scale error
- <1%From approximately 50–70%
- Reported fit
- ≈0.99R² · alongside MAE and RMSE
Interactive companion
A small experiment
you can feel.
Explore controlled contact, a magnetic field, and a network learning from touch. This is an educational model, separate from the professional research system.