Echo State Networks for Estimating Exteroceptive Conditions From Proprioceptive States in Quadruped Robots

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Proprioceptive sensors provide local joint data in robots.
• Exteroceptive sensors are typically needed for global environmental information.
• Current methods often require complex sensor fusion or specific robot designs.

Purpose of the Study:

• To develop a unified methodology for robots to infer global information from local proprioceptive data.
• To enable concurrent dynamic estimation and terrain classification using a single computational structure.
• To create a soft sensor approach applicable to various robotic platforms.

Main Methods:

• Utilizing reservoir computing, a form of recurrent neural network, to process proprioceptive leg joint data.
• Mapping local joint kinematics to exteroceptive information like ground reaction forces and terrain type.
• Implementing the reservoir computing structure as a soft sensor device within a simulated and real quadruped robot.

Main Results:

• The reservoir computing model successfully estimated ground reaction forces and classified terrain types concurrently.
• The methodology demonstrated robustness across different terrains and simulated sensory system faults.
• The approach proved independent of the specific robotic mechanical design.

Conclusions:

• Reservoir computing offers an effective data-driven strategy for enhancing robot environmental perception.
• The proposed soft sensor method generalizes well to different robotic structures and conditions.
• This approach reduces reliance on complex exteroceptive sensor suites.