Incrementally learning objects by touch: online discriminative and generative models for tactile-based recognition.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Humans excel at object recognition via touch, improving with experience.
• Robotic tactile perception is crucial for interaction and manipulation.
• Incremental learning allows systems to adapt and improve over time.

Purpose of the Study:

• To develop and compare novel discriminative and generative tactile learners for object recognition.
• To enable iterative improvement in tactile recognition through online learning.
• To implement incremental unsupervised learning mechanisms for situations lacking teacher labels.

Main Methods:

• Utilized the sparse online infinite Echo-State Gaussian process (OIESGP) framework.
• Developed two novel online learners for discriminative and generative tactile recognition.
• Incorporated incremental unsupervised learning using novelty scores and extreme value theory.

Main Results:

• Classifiers performed comparably to state-of-the-art C4.5 and SVM methods.
• Accurate object classifications were achieved using only 20-30% of the grasp sequence.
• Unsupervised learning methods produced high-quality clusterings, outperforming sequential k-means and SOM.

Conclusions:

• Tactile signals are highly relevant for accurate object classification in robots.
• Incremental learning significantly enhances robotic tactile recognition capabilities.
• Proposed methods offer effective solutions for both supervised and unsupervised tactile learning tasks.