Stiffness Mapping of Deformable Objects through Supervised Embedding and Gaussian Process Regression

This Master’s thesis presents a comprehensive approach to enabling robots to understand and map the stiffness properties of deformable objects through advanced machine learning techniques. The work addresses a fundamental challenge in robotics: how to enable machines to perceive and quantify the mechanical properties of soft, deformable materials through tactile interaction.

Research Problem: Traditional robotic systems struggle with handling deformable objects because they lack the ability to perceive and adapt to varying stiffness properties across different regions of soft materials. This limitation significantly restricts their application in domains such as medical robotics, food handling, and manufacturing of soft goods.

Technical Approach: The thesis combines two powerful machine learning methodologies:

• Supervised Embedding: Develops learned representations that capture the essential characteristics of tactile interactions with deformable objects
• Gaussian Process Regression: Provides a probabilistic framework for predicting stiffness properties across unobserved regions of objects, including uncertainty quantification

Key Contributions:

• Novel integration of supervised embedding techniques for tactile data representation
• Application of Gaussian process regression for spatial stiffness prediction with uncertainty estimates
• Comprehensive experimental validation demonstrating the approach’s effectiveness on various deformable materials
• Framework that enables continuous learning and adaptation as more tactile data is collected

Significance: This work represents an important step toward enabling robots to develop sophisticated understanding of material properties through touch. The combination of embedding techniques and probabilistic regression provides both accurate predictions and uncertainty estimates, which are crucial for safe robotic manipulation of deformable objects.

The research has broad implications for applications requiring careful handling of soft materials, including surgical robotics, food processing, and assistive technologies. The probabilistic nature of the approach also makes it well-suited for safety-critical applications where understanding prediction confidence is essential.

Completed at Carnegie Mellon University’s Robotics Institute under the guidance of faculty experts in robotic manipulation and machine learning.

Recommended citation: E. Harber. (2022). "Stiffness Mapping of Deformable Objects through Supervised Embedding and Gaussian Process Regression." Master's thesis, Carnegie Mellon University, Pittsburgh, PA.
Download Paper