Asymptotically Optimal Motion Planning for Learned Tasks Using Time-Dependent Cost Maps.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robots operating in unstructured environments require sophisticated motion planning to avoid collisions and satisfy task-specific constraints.
• Existing methods often struggle with complex tasks that involve maintaining specific object orientations or trajectories.

Purpose of the Study:

• To develop a sampling-based motion planning method that enables robots to avoid obstacles while adhering to learned task motion constraints.
• To enable intuitive robot training through demonstrations for autonomous operation in cluttered, real-world scenarios.

Main Methods:

• A time-dependent cost metric is learned from demonstrations to capture essential task motion features.
• A sampling-based motion planner utilizes the learned cost metric for collision-free path computation.
• The planner is asymptotically optimal, minimizing Mahalanobis distance to demonstration distributions and reducing computation time.

Main Results:

• The method successfully computes motion plans that simultaneously avoid obstacles and satisfy learned task constraints.
• Demonstrated effectiveness and speed on a humanoid robot performing tasks like transferring powder and pushing buttons.
• The planner leverages demonstration data to significantly decrease computation time.

Conclusions:

• The proposed motion planning approach enables robots to autonomously perform tasks in cluttered environments by learning from demonstrations.
• This method enhances robot adaptability and efficiency in handling unforeseen obstacles and complex task requirements.
• Future work aims to extend this approach to more intricate robotic tasks.