Soft DAgger: Sample-Efficient Imitation Learning for Control of Soft Robots.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Soft robots offer unique advantages in dexterity and safety but present control challenges.
• Traditional control methods often struggle with the high dimensionality and nonlinear dynamics of soft robots.
• Imitation learning (IL) offers a promising avenue for training robot controllers by learning from demonstrations.

Purpose of the Study:

• To introduce Soft DAgger, an efficient imitation learning algorithm for training soft robot controllers.
• To demonstrate the algorithm's effectiveness on a soft robotic arm performing a 3D writing task.
• To achieve robust control and generalization without relying on extensive exploration or reinforcement learning.

Main Methods:

• Soft DAgger utilizes a dynamic behavioral map to translate the robot's task space to its actuation space.
• The algorithm learns from expert demonstrations, state-action history, and the robot's current position.
• Two variants of the control algorithm were proposed and evaluated.

Main Results:

• The proposed algorithm enabled a soft robotic arm to perform 3D letter writing with good generalization.
• Soft DAgger achieved improved task reproducibility and a significant decrease in optimization time and samples.
• The approach demonstrated effective control without costly exploration or reinforcement learning.

Conclusions:

• Soft DAgger provides a practical and sample-efficient solution for controlling soft robots in complex tasks.
• This study represents an initial exploration of imitation learning with online optimization for soft robot control.
• The algorithm shows potential for advancing the capabilities and applications of soft robotic systems.