Nature and evolution have produced remarkably capable organisms, inspiring engineers and researchers to create robots that replicate the versatility and resilience from the highly complex and functional physiologies of the natural. As a result, soft and bio-inspired robotics have emerged as a promising area of innovation in recent years as their unprecedented flexibility and adaptability, but their continuously deformable structures pose significant challenges for modeling and control. Machine learning has emerged as a powerful tool for soft robot perception, enabling proprioception and control through sensor-driven neural networks. However, these approaches face two key challenges: performance degradation over time due to material wear and sensor shifts, and the difficulty of learning meaningful representations in highly underactuated systems as onboard sensors may not provide a signal informative enough.

Our research systematically investigates how neural network performance deteriorates as system parameters change, using soft robotic simulations of a trunk and a cable-actuated gripper. We additionally propose a novel latent vision-aided sensor recalibration framework that utilizes vision-based latent encodings to fine-tune sensor models without requiring full retraining or continuous visual input during deployment. This approach retains the benefits of vision systems while mitigating their computational and environmental limitations, improving the adaptability of learning-based soft robot models and reducing the need for costly and time-intensive retraining.

I am participating in the SuperUROP program because I am interested in exploring the intersection between machine learning and robotics. I am excited to be able to develop additional research skills as well as pursue a more independent project. I have previously taken subjects in the domain and performed machine learning research, and so I hope to leverage my knowledge towards real-world applications.