Adaptive and predictive control of a simulated robot arm.

Area of Science:

• Robotics
• Computational Neuroscience
• Machine Learning

Background:

• Traditional robot control struggles with motor and distal error problems.
• Sensor data (position, velocity, acceleration) can be leveraged for control.
• Bio-inspired cerebellar circuitry offers potential for robust motor learning.

Purpose of the Study:

• To develop a robot control system that learns motor control from sensor error estimates without explicit motor error knowledge.
• To integrate a cerebellar neural layer with a machine learning engine for adaptive control.
• To evaluate the scalability of the proposed control scheme for high degrees of freedom (DOFs) robot arms.

Main Methods:

• A recurrent loop embedding a cerebellar neural layer and a machine learning engine (locally weighted projection regression - LWPR).
• Input decomposition to facilitate learning using LWPR for incremental forward model learning.
• An adaptive feedback (AF) controller for precise, compliant, and stable object manipulation.

Main Results:

• The cerebellar-LWPR synergy enables robots to adapt to changing conditions.
• The system learns motor control effectively using available sensor error estimates.
• Demonstrated precise, compliant, and stable control during object manipulation tasks.

Conclusions:

• The integration of cerebellar circuitry and LWPR provides an efficient and adaptable robot control architecture.
• This bio-inspired machine learning approach overcomes limitations of traditional robot control methods.
• The proposed scheme shows promise for controlling complex, high-DOF robot arms like light weight robots (LWRs).