Bidirectional Tutoring for Developmental Motor Learning in Robots: Co-Developed Interaction Dynamics Support Stable Learning

This paper investigates how the nature of tutor-robot interaction—specifically whether it is unidirectional or bidirectional—affects the stability and generalization of motor skill acquisition. While traditional imitation learning treats tutoring as a passive, one-way demonstration, the authors hypothesize that bidirectional interaction acts as a scaffold, allowing the robot's prior knowledge to shape new learning and preserve behavioral coherence across developmental stages.

The researchers implemented a developmental learning framework using a Predictive coding-inspired Variational Recurrent Neural Network (PV-RNN) extended with generative replay to mitigate catastrophic forgetting. They compared two tutoring modes:

• Unidirectional Tutoring: The robot is passive (gravity compensation mode) and follows the tutor's demonstration.
• Bidirectional Tutoring: The robot actively generates actions based on its current internal model, and the tutor intervenes only when necessary to correct the behavior.

The study tested these modes using a physical humanoid robot (Torobo) in an object manipulation task. The experiments were conducted in two settings: one with a human tutor and one with an AI tutor, ensuring that the observed effects were robust across different intervention styles.

Bidirectional tutoring consistently outperformed unidirectional tutoring in fostering stable motor learning. By coupling tutor intervention with the robot's ongoing, self-generated behavior, the system developed more coherent sensorimotor representations. This coherence allowed the robot to generalize better to new object positions and required less tutor guidance over time. The results suggest that treating motor learning as a co-developed, socially grounded process is more effective than treating it as a passive data-collection exercise.