← Engineering work

ML Engineering · SERES · June–August 2026

Tactile sensing AI
& automated testing.

Magnetic tactile sensing
& machine learning

I led tactile ML and model development to improve how a robot finger estimates contact position and force. I also designed the test bed and built the software-side force control, automated tests, and analysis workflow. My work included reference-point selection, spatial-resolution methodology, test design, and calibration.

Role
ML Engineering
SERES · June–August 2026
My contribution
AI development, test-bed design, and automation
Characterization, calibration, and test methodology
Outcome
Improved contact estimates
Programmable, repeatable experiments

My contributions

Test-bed design,
data collection & ML.

Experiment

Designing a force-controlled test bed

I designed a new test bed to replace a limited, inaccurate setup and make contact position and load repeatable. I developed the test methodology and test matrix to structure those experiments.

The Torbal setup was not force-aware. My software added force control, letting engineers specify target loads and run automated multi-step tests.

Measurement

Developing sensor characterization and calibration methods

I researched and developed methods for choosing the number of force reference points and contact locations, determining spatial resolution, and calibrating the system.

I also integrated automated data collection, graphing, and analysis into the testing software, connecting sensor readings with contact position and applied force for model training and evaluation.

Learning

Predicting contact position and force

I led tactile ML and model development, improving the previous model’s ability to estimate contact position and force from magnetic measurements.

The two-headed network uses a shared learned representation, then separates into position and force predictions.

Evaluation

Evaluating the model on held-out data

Evaluation used two training runs and a separate holdout run, with millions of measurements at each stage. MAE, RMSE, and model fit helped assess the predictions.

The improved model’s reported results include full-scale error below 1% and R² around 0.99 within the research setup.

Engineering outcome

Model accuracy
& repeatable testing.

The resulting platform connects controlled loading, sensor measurements, and model evaluation. I documented the software and methods so the experiments can be repeated and extended.

The previous system had reported MAE of up to 70% and RMSE of up to 100%, alongside a limited, inaccurate test bed.

Reported full-scale error
<1%Improved model · research setup
Reported fit
≈0.99R² · research setup

Interactive learning lab

Explore contact and force prediction.

An educational simulation of contact, magnetic sensing, and learning, using synthetic data.

INTERACTIVE TACTILE LABSYNTHETIC SENSOR MODEL

Train a tactile-sensing model.

Magnetic tactile sensing: contact, magnetic displacement, and measurementA cutaway diagram of a compliant sensing surface. A probe indents the elastomer, shifting embedded magnets relative to a fixed sensor board. Magnetic measurements feed the contact-estimation model. This explains the principle and is not a measured research result.APPLIED FORCEELASTOMERMAGNETSSENSORSINFERENCE

Move the contact. Change the force. Watch a small network learn the relationship.

REFERENCE FORCE—
INFERRED FORCE—
POSITION RMSE—
0 / 3,000TRAINING UPDATES

Instrument controls

Move the contact on the fingertip, or use the X and Y controls. All values here are normalized, synthetic measurements.

Change the object’s compliance: the same travel produces a different force. Force feedback adjusts the travel to meet the target.

Reference X / Y
— / —
Actual actuator travel
—
Force target error
—
Predicted X / Y
— / —
Force MAE / RMSE
— / —
Clean held-out MSE
—
Training / held-out examples
1,536 / 512
Network
12 → 24 → 24 → 3 · 987 parameters

Validation MAE and RMSE use the same clean held-out examples throughout training. Added noise affects the current sensor readings and inference, not this fixed benchmark.

123456789101112

Synthetic demonstration of the sensing principle. This small browser model uses generated data; it is not the SERES sensor, dataset, or research network. Related magnetic sensing research ↗

Synthetic sensor demonstration · normalized units. This browser network and generated data illustrate the principle; they are not the SERES research model or dataset.

Paused. Controls can still inspect a single state.