Learning-Based Control for Soft Robot-Environment Interaction with Force/Position Tracking Capability.

Area of Science:

• Robotics
• Control Theory
• Machine Learning

Background:

• Soft robots offer safe human-environment interaction but face control challenges due to complex dynamics and unknown environments.
• Developing analytical models for soft robots is difficult, hindering effective control strategies.

Purpose of the Study:

• To propose a learning-based optimal control approach for soft robot-environment interaction.
• To address challenges posed by nonlinear dynamics, unknown environments, and uncertainties.

Main Methods:

• An optimized combination of a feedforward controller (probabilistic model predictive control) and a feedback controller (nonparametric learning methods).
• A purely data-driven approach requiring no prior knowledge of robot dynamics or environment structures.
• Online updating capability for adaptation to unknown environments.

Main Results:

• The proposed approach demonstrated stability and convergence through theoretical analysis.
• A soft robotic manipulator successfully tracked target positions and forces during interactions with a manikin.
• Outperformed other data-driven control methods in comparative tests.

Conclusions:

• The work presents a viable learning-based control approach for soft robot-environment interactions.
• The method provides robust force and position tracking capabilities for soft robotic systems.