insideOut: A Bio-Inspired Machine Learning Approach to Estimating Posture in Robots Driven by Compliant Tendons

Area of Science:

• Robotics
• Machine Learning
• Biomechanics

Background:

• Accurate limb posture estimation is vital for robotic control.
• Traditional joint angle sensors complicate mechanical design in compliant tendon-driven systems.
• Nonlinear tendon stiffness in these systems decouples motor and joint angles, hindering direct estimation.

Purpose of the Study:

• To develop a novel machine learning algorithm for accurate joint posture estimation in dynamic, tendon-driven robotic systems.
• To investigate the efficacy of using non-collocated sensory information, including motor angles and tendon tensions, for posture estimation.
• To compare the performance of artificial neural networks (ANNs) trained with different sensory inputs.

Main Methods:

• A simulation of an inverted pendulum driven by motors and nonlinear tendons was used.
• Artificial neural networks (ANNs) were trained using motor angles and tendon tensions.
• ANNs were also trained using only motor angles for comparison.
• Performance was evaluated using cross-validation with novel movements.

Main Results:

• ANNs trained with both motor angles and tendon tensions demonstrated significantly higher accuracy in predicting joint angles.
• The inclusion of tendon tension data improved estimation accuracy regardless of tendon properties or mechanical hyperparameters.
• Results were robust across various network and mechanical configurations.

Conclusions:

• Tendon tension is a crucial sensory signal for improving joint angle estimation in compliant, tendon-driven robotic systems.
• Machine learning, specifically ANNs utilizing non-collocated sensory data, offers a viable solution for complex robotic posture control.
• This approach mimics biological mechanoreception, suggesting potential for bio-inspired robotic design.