Unsupervised Sim-to-Real Adaptation of Soft Robot Proprioception Using a Dual Cross-Modal Autoencoder.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Data-driven calibration enhances proprioception in soft robotics.
• Numerical simulation offers computational efficiency but suffers from domain gaps.
• Accurate, generalized application is limited by discrepancies between simulated and real sensor data.

Purpose of the Study:

• To propose an unsupervised domain adaptation framework for aligning heterogeneous sensor domains in soft robotics.
• To enable data-efficient and generalized calibration by bridging the simulation-reality gap.
• To integrate domain adaptation with anomaly detection for enhanced robot perception.

Main Methods:

• Developed a dual cross-modal autoencoder for feature-level domain matching without extensive labeling.
• Integrated unsupervised domain adaptation with anomaly detection for collision detection.
• Applied the framework to a multigait soft robot for shape estimation and collision detection tasks.

Main Results:

• Achieved accurate digital-twinned calibration in both simulated and real-world domains.
• Demonstrated superior performance compared to existing benchmarks for shape estimation and collision detection.
• Validated the framework's transferability and data efficiency across different tasks.

Conclusions:

• The unsupervised domain adaptation framework offers a novel approach for embodied intelligence in soft robotics.
• The methodology effectively aligns heterogeneous sensor domains, improving perception capabilities.
• This work paves the way for more robust and adaptable soft robotic systems by blending simulation and real-world data.