Curriculum-based humanoid robot identification using large-scale human motion database.

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Accurate dynamics models are crucial for humanoid robot motion control and planning.
• Existing methods face challenges with unreliable initial models and complex optimization for parameter identification.
• Data acquisition for model identification requires feasible motions and efficient movement design.

Purpose of the Study:

• To develop a humanoid robot dynamics model identification method that bypasses the need for a good initial model.
• To create a method that avoids solving highly nonlinear optimization problems for movement design.
• To improve the efficiency and accuracy of humanoid robot dynamics model identification.

Main Methods:

• Proposed a curriculum learning approach for gradual identification of dynamics models from unreliable initial models.
• Utilized a large-scale human motion database to efficiently design humanoid movements for parameter identification.
• Evaluated the method using simulation experiments on an 18-Degree-of-Freedom (DoF) simulated upper-body humanoid robot.

Main Results:

• The curriculum-based approach demonstrated more efficient identification of humanoid model parameters compared to random motion selection.
• The method successfully acquired a wide variety of motion data, leading to efficient parameter estimation.
• An accurate dynamics model of the simulated humanoid robot was successfully identified.

Conclusions:

• The proposed curriculum learning strategy effectively addresses the challenges of initial model unreliability and complex optimization in humanoid robot dynamics identification.
• This method offers a more efficient and practical approach to obtaining accurate dynamics models for humanoid robots.
• The findings contribute to advancing the capabilities of humanoid robot control and motion planning.