Calibration of Visuo-based Tactile Sensor

Role: Research Fellow & Programmer
Timeline: 4 Weeks (Academic Research Fellowship, Two-Person Team)
Location: King's College London
Tech Stack: [Python] [TensorFlow / PyTorch] [Data Analysis]

The Context

Tactile sensors capture the deformation of a contact surface to develop an "artificial sense of touch." However, due to the varying shapes of contact surfaces (e.g., a screwdriver vs. a ball), there is no simple linear relationship between the force applied and the resulting deformation. This research project aimed to build an algorithm capable of learning this complex relationship to accurately predict contact forces.

My Contributions

• Deep Learning Architecture: Researched and developed a Convolutional Neural Network (CNN) from scratch to process visual deformation data and map it to force values.
• Data Engineering: Planned and executed the data collection procedure, capturing thousands of deformation images mapped to ground-truth forces using a Nano 17 force sensor and a 6-axis force plate.
• Model Optimization: Trained and fine-tuned the network independently, successfully achieving a prediction accuracy within 0.1N of the ground-truth force values across various deformation types.

The Outcome

Delivered the finalized neural network ahead of schedule and published a comprehensive 16-page technical report detailing the mathematical correlation and calibration effectiveness.