A User Study on Robot Skill Learning Without a Cost Function: Optimization of Dynamic Movement Primitives via Naive...

Area of Science:

• Robotics
• Human-Robot Interaction
• Machine Learning

Background:

• Teaching robots new skills is a significant challenge in personal robotics.
• Current methods often require complex cost functions and external systems.
• Dynamic Movement Primitives (DMPs) offer a promising approach for skill learning.

Purpose of the Study:

• To investigate the effectiveness of user-provided discrete feedback for robot skill learning.
• To compare learning performance using user feedback versus objective cost functions.
• To assess the ability of non-expert users to guide robot learning.

Main Methods:

• A user study involving participants teaching the robot Pepper a game of skill.
• Utilizing a state-of-the-art skill learning method based on dynamic movement primitives (DMPs).
• Comparing learning with discrete user ratings against an objectively determined cost function.

Main Results:

• An intuitive graphical user interface for discrete feedback enables robot learning of complex movement skills.
• Non-expert users naturally employ effective rating strategies, achieving performance comparable to objective cost functions.
• DMP-based optimization makes tedious cost function definition obsolete.

Conclusions:

• Discrete user feedback is a viable and effective method for robot skill acquisition.
• Robot learning can be successfully guided by non-expert users without prior knowledge of the algorithms.
• Further research can improve continuous movement skill learning from human input.