A Self-Supervised Learning Framework for Soft Robot Proprioception.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Soft robots offer safety and adaptability advantages but present proprioception challenges.
• Machine learning, particularly supervised learning (SL), has been used for soft robot proprioception.
• SL methods require extensive annotated data, hindering real-world application.

Purpose of the Study:

• To develop a self-supervised learning (SSL) framework for enhanced soft robot proprioception.
• To reduce the reliance on costly annotated data for training soft robot control systems.

Main Methods:

• Proposed a novel self-SL framework for soft robot proprioception.
• Utilized vast unannotated data for initial network pretraining via self-SL.
• Fine-tuned the pretrained model using a limited set of annotated samples via SL.

Main Results:

• Validated the framework on a 3-D morphological reconstruction task using public data.
• Achieved superior performance compared to fully supervised methods.
• Required only approximately 1/20th of the annotated samples needed by traditional SL.

Conclusions:

• The proposed self-SL framework significantly improves the efficiency of training soft robot proprioception models.
• This approach mitigates the data annotation bottleneck, enabling faster adoption of soft robots in real-world scenarios.