Learning to Perform Trajectory Generation From Low-Quality Demonstrations.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Human-robot skill transfer is crucial for robot learning, enabling robots to acquire new skills through human guidance.
• Current methods often struggle with imperfect human demonstrations due to errors, sensor issues, or data variations.
• These imperfections lead to challenges like unrelated data, information loss, and inconsistent demonstration lengths and amplitudes.

Purpose of the Study:

• To propose a new trajectory alignment and filtering method for extracting useful information from multiple, potentially imperfect, human demonstrations.
• To enhance the robustness of learning from demonstrations (LfD) by addressing variations and errors in training data.
• To enable robots to learn and generate skills effectively despite inconsistencies in human demonstrations.

Main Methods:

• Developed a novel trajectory alignment and filtering technique to preprocess demonstration data.
• Integrated this method with probabilistic movement primitives (ProMPs) as a case study for probabilistic movement learning.
• Applied the method to extract relevant features from multiple demonstrations of varying quality.

Main Results:

• The proposed method effectively extracts relatively useful information from multiple demonstrations, even with imperfections.
• Simulation results verified the effectiveness of the trajectory alignment and filtering approach.
• The technique allows robots to learn and generate trajectories for skill completion from varied quality demonstrations.

Conclusions:

• The developed trajectory alignment and filtering method significantly improves the quality of data used in learning from demonstrations (LfD).
• This approach enhances the robot's ability to learn skills accurately from imperfect human demonstrations.
• The method is compatible with various probabilistic movement learning techniques, offering broad applicability in robotics.