Deep Learning-Based Contact Force Control for a Robotic Leg.

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Stable contact force control is crucial for humanoid robot balance during bipedal locomotion.
• Admittance controllers, commonly used for this purpose, struggle with the inherent nonlinearities of robot systems.

Purpose of the Study:

• To develop a learning-based contact force controller that overcomes the limitations of traditional admittance controllers.
• To improve the stability and performance of force control in humanoid robots.

Main Methods:

• A deep neural network (DNN) was employed as an inverse model to capture system nonlinearities using input-output data.
• A proportional-integral (PI) controller was integrated to minimize steady-state errors.
• The controller computes target foot height based on force and height measurements, without needing a dynamic robot model.

Main Results:

• The proposed DNN-based controller demonstrated significant performance improvements over an admittance controller.
• Overshoot was reduced by an average of 96% and settling time by 61% in step responses.
• Root-mean-square error (RMSE) for force tracking decreased by an average of 66.3% across step and sinusoidal experiments.

Conclusions:

• The learning-based contact force controller effectively enhances stability and reduces errors in humanoid robot locomotion.
• This approach offers a robust alternative to traditional methods for controlling contact forces in complex robotic systems.