Human-robot skills transfer interfaces for a flexible surgical robot

Area of Science:

• Robotics
• Surgical Technology
• Machine Learning

Background:

• Current robot-assisted surgical systems face limitations due to rigid tools.
• Minimally invasive surgery requires flexible tools for navigating narrow openings and manipulating soft organs.
• The STIFF-FLOP project aims to develop a soft robotic arm for enhanced surgical capabilities.

Purpose of the Study:

• To design learning interfaces for transferring skills from human demonstrations to soft robotic arms.
• To enable robot programming by demonstration, focusing on imitating underlying intent rather than simple action mimicry.
• To develop a method for extracting an objective function that explains demonstrations for skill self-refinement.

Main Methods:

• Utilizing context-dependent reward-weighted learning to identify relevant objective functions based on task phase and situation.
• Extracting an objective function from an over-specified set of candidate reward functions.
• Testing the approach in simulation using a cutting task performed by the STIFF-FLOP flexible robot with kinesthetic demonstrations.

Main Results:

• The proposed approach enables the robot to learn the relevance of candidate objective functions.
• The robot utilizes learned objective functions for skill refinement in the policy parameters space.
• Successful simulation of a cutting task using the STIFF-FLOP robot and human demonstrations.

Conclusions:

• The developed learning interface facilitates skill transfer from human demonstration to soft robotic systems.
• Context-dependent reward-weighted learning allows robots to adapt and refine skills based on task context.
• This approach advances the capabilities of soft robotic arms for complex surgical procedures.