An online learning algorithm for adapting leg stiffness and stride angle for efficient quadruped robot trotting

Area of Science:

• Robotics
• Biomechanical Engineering
• Machine Learning

Background:

• Animals exhibit adaptive leg control for stability and energy efficiency on varied terrains.
• Legged robots often require precise terrain knowledge and pre-programmed gaits for optimal performance.
• Current control methods for legged robots can be computationally intensive and require extensive training.

Purpose of the Study:

• To develop an online learning algorithm that enables legged robots to adapt leg stiffness and stride angle in real-time.
• To minimize the cost of transport for legged robots on unknown ground conditions without prior knowledge.
• To emulate the energy efficiency and stability observed in animal locomotion.

Main Methods:

• An approximate stochastic gradient method is employed for real-time parameter adaptation.
• The algorithm is model-free, requiring no precise robot model for operation.
• Legged robot control parameters (stiffness, stride angle) are adjusted adaptively during traversal.

Main Results:

• The algorithm demonstrated computational efficiency and suitability for real-time robotic applications.
• Simulations and experiments on a quadruped robot showed convergence to near-optimal cost of transport values.
• The system achieved improved performance without prior knowledge of terrain or gait conditions.

Conclusions:

• The proposed online learning algorithm effectively enhances energy efficiency in legged robots.
• The model-free and computationally efficient nature makes it broadly applicable to various legged robots.
• This approach offers a robust method for adaptive control of compliant legged robots.