A Framework for Sensorimotor Cross-Perception and Cross-Behavior Knowledge Transfer for Object Categorization.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Humans intuitively understand object properties through exploratory behaviors.
• Robots can infer physical properties using exploratory actions beyond visual data.
• Current robot learning is robot-specific and lacks cross-modal or cross-robot transferability.

Purpose of the Study:

• To develop a framework for knowledge transfer across different robotic behaviors and sensory modalities.
• To enable robots to share learned sensorimotor knowledge.
• To address limitations of robot-specific models and sensor failures.

Main Methods:

• Proposed a framework for knowledge transfer using variational auto-encoders and encoder-decoder networks.
• Hypothesized that shared multi-sensory object observations enable mappings between feature spaces.
• Utilized a dataset with a robot performing 9 behaviors and 4 sensory modalities on 100 objects.

Main Results:

• Demonstrated successful transfer of sensorimotor knowledge across behaviors.
• Showcased effective knowledge transfer across different sensory modalities.
• Validated the framework's ability to bootstrap recognition models on new robots or sensor configurations.

Conclusions:

• Sensorimotor knowledge about objects is transferable across behaviors and sensory modalities.
• The proposed framework allows robots to leverage existing knowledge for faster learning.
• This facilitates more adaptable and robust robotic systems capable of handling sensor limitations.