Task Feasibility Maximization Using Model-Free Policy Search and Model-Based Whole-Body Control.

Area of Science:

• Robotics
• Control Theory
• Artificial Intelligence

Background:

• Generating feasible motions for highly redundant robots, like humanoids, is a complex, high-dimensional challenge.
• Model-based whole-body control can create dynamic behaviors but often struggles with task feasibility due to isolated planning.
• Existing methods lack guarantees for motion accomplishment, necessitating improved approaches.

Purpose of the Study:

• To develop a novel optimization loop for enhancing task feasibility in robot motion generation.
• To address the intractability of complex motion planning problems using combined control strategies.
• To demonstrate the practical realization of initially infeasible dynamic motions on humanoid robots.

Main Methods:

• Implemented a proof-of-concept optimization loop integrating model-free policy search with model-based whole-body control.
• Utilized policy search to automatically refine task parameters for improved feasibility.
• Validated the approach through experiments on both simulated and real iCub humanoid robots.

Main Results:

• Successfully optimized task feasibility, enabling the execution of complex dynamic motions.
• Demonstrated the realization of a challenging sit-to-stand transition on the iCub humanoid robot.
• Showcased the synergistic benefits of combining model-free and model-based control methods.

Conclusions:

• The proposed optimization loop effectively improves the feasibility of robot motion planning.
• This integrated approach overcomes limitations of using either control strategy alone.
• The method holds promise for enabling more complex and reliable dynamic behaviors in humanoid robots.