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 limited experience.
• Large language models (LLMs) require vast datasets (billions of tokens) for training.
• The mechanisms underlying efficient human developmental learning remain an open question.

Purpose of the Study:

• To investigate the computational principles of efficient, intrinsically motivated language acquisition.
• To model human-like developmental learning in robotic agents using curiosity-driven exploration.
• To compare simulated learning patterns with observations from child language development.

Main Methods:

• Utilized curiosity-driven self-exploration in robotic agents to learn action-sentence associations.
• Employed active inference amortized with Q-learning for intrinsically motivated developmental learning.
• Simulated learning processes to analyze generalization, learning speed, and developmental patterns.

Main Results:

• Compositional generalization improved significantly with increased scale of elements.
• Curiosity-driven exploration accelerated the learning process.
• Rote action-sentence pairing was observed to precede compositional generalization.
• Exception handling led to U-shaped performance curves, mirroring representational redescription in child language.

Conclusions:

• Curiosity-driven active inference provides a framework for intrinsically motivated sensorimotor-linguistic learning.
• This approach supports scalable compositional generalization and exception handling in artificial agents.
• Findings suggest active inference may explain efficient developmental learning in humans and robots.