Active learning strategies for robotic tactile texture recognition tasks.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Robots require accurate texture classification for environmental perception and decision-making.
• Vast amounts of sensor time-series data pose challenges for robot learning.
• Human feedback is crucial for robot learning, especially when misclassifications occur.

Purpose of the Study:

• To reduce human labeling effort in robot texture classification using Active Learning (AL).
• To develop a novel class-balancing instance selection algorithm for AL strategies.
• To evaluate the impact of sliding window sizes on AL performance for texture classification.

Main Methods:

• Implemented Active Learning (AL) to select informative samples for annotation.
• Utilized a sliding window strategy for time-series feature extraction.
• Proposed and integrated a class-balancing instance selection algorithm with standard AL.

Main Results:

• AL strategies reduced training data by up to 70% while maintaining or surpassing baseline performance.
• The novel class-balancing algorithm improved AL performance in multi-class texture datasets.
• A 6-second sliding window achieved the best performance, with Extra Trees yielding an average f1-score of 90.21%.

Conclusions:

• The proposed data pipeline enhances texture classification performance for robots.
• Active Learning, combined with class-balancing, significantly reduces the need for human labeling.
• Optimized window size and model selection are key for high-performance robotic texture classification.