Guided Stochastic Optimization for Motion Planning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Learning from Demonstration (LfD) methods assist robots with complex manipulation tasks.
• Current LfD approaches are often control-based, limiting their adaptability and platform compatibility.
• Task execution and learning scalability remain challenges in collaborative robotics.

Purpose of the Study:

• To propose a novel motion planning approach integrating LfD with generic motion planning for enhanced robustness and scalability.
• To enable skill transfer between different robot platforms.
• To improve task success rates and adaptability to changing environments.

Main Methods:

• Introduced Dynamical Movement Primitives (DMPs) based LfD as initial trajectories for the Stochastic Optimization for Motion Planning (STOMP) framework.
• Developed Guided Stochastic Optimization for Motion Planning (GSTOMP).
• Evaluated GSTOMP on two different manipulator systems in simulation and real-world conditions.

Main Results:

• GSTOMP demonstrated improved task success rates compared to simple LfD methods.
• Successful skill transfer between dissimilar robot platforms was achieved with good performance.
• GSTOMP's motion planning performance was comparable or superior to existing state-of-the-art motion planners.

Conclusions:

• The proposed GSTOMP approach offers a robust and scalable solution for robot task learning and execution.
• GSTOMP effectively addresses limitations of traditional LfD methods by integrating generic motion planning.
• The framework facilitates efficient skill transfer, enhancing robot versatility across different platforms.