Curiosity-driven development of action and language in robots through self-exploration

Area of Science:

• Cognitive Science
• Developmental Psychology
• Robotics

Background:

• Human infants exhibit remarkable language generalization from minimal experience.
• Large language models require extensive training data, contrasting with human efficiency.

Purpose of the Study:

• Investigate the mechanisms underlying efficient human language development.
• Explore how curiosity-driven self-exploration facilitates learning in artificial agents.

Main Methods:

• Robotic agents learned actions from imperative sentences using curiosity-driven self-exploration.
• Active inference was amortized using Q-learning for intrinsically motivated learning.

Main Results:

• Generalization improved with increased compositional elements.
• Curiosity-driven exploration accelerated learning.
• Rote learning preceded compositional generalization.
• Exception handling led to U-shaped performance, similar to human representational redescription.

Conclusions:

• Curiosity-driven active inference supports scalable compositional generalization and exception handling.
• This model accounts for intrinsically motivated sensorimotor-linguistic learning in humans and AI.