Online sensorimotor learning and adaptation for inverse dynamics control.

Area of Science:

• Robotics
• Neuroscience
• Machine Learning

Background:

• Human arm control exhibits remarkable efficiency and adaptability.
• Current robotic systems often require extensive data for learning complex movements.
• Bridging the gap between biological and artificial sensorimotor control remains a challenge.

Purpose of the Study:

• To develop a data-efficient model for human-like arm inverse dynamics control.
• To investigate a novel approach combining feedforward and adaptive feedback mechanisms.
• To enable rapid learning and adaptation in robotic arm movements.

Main Methods:

• Proposed the sensorimotor learning and adaptation (SEED) model.
• Utilized a feedforward Gaussian motor primitive (GATE) neural network.
• Incorporated an adaptive feedback impedance (AIM) mechanism for online gain tuning.

Main Results:

• The SEED model achieved stable task learning within 10 trials on a two-joint robot arm.
• Demonstrated significantly faster learning compared to state-of-the-art deep learning methods requiring thousands of trials.
• Revealed that the elbow joint requires minimal active control (<3% effort), aligning with the proximal-distal control gradient in human movement.

Conclusions:

• The SEED model offers a data-efficient solution for sensorimotor learning and adaptation.
• This approach facilitates exploration of unknown dynamics and human-like control strategies.
• Paves the way for implementing efficient, human-like robotic arm movements.