Learning Dominant Dynamics for Continuum Robot Control (John Alora, PhD Defense)

John Alora PhD Defense (12/17/2024)

Continuum robotics, inspired by the fluidity of living systems, offers transformative potential across applications from ocean exploration to enabling space and planetary missions. However, the compliance and intricate motion of continuum robots pose significant challenges for modeling and real-time control due to their high-dimensional, nonlinear dynamics. This research contributes flexible and computationally efficient methods that offer a promising foundation for safely deploying continuum robots in high-stakes, resource-constrained environments. We first introduce a model-based control methodology based on Spectral Submanifold (SSM) Reduction. This approach leverages SSMs as structural priors for data-driven system identification. By doing this, we learn the dominant dynamics of the underlying system and exploit its low dimensionality for efficient model predictive control (MPC). To guarantee safety in the face of model uncertainty, we leverage the SSM structure to quantify compression error and design a robust tube MPC that ensures constraint satisfaction. We then extend our modeling approach to handle more dynamic tasks by learning time-varying SSMs via latent space interpolation. Finally, we evaluate our approach for different robot embodiments in high-fidelity simulation and on a diamond robot platform, thoroughly benchmarking against state-of-the-art methods in the continuum robot literature. While demonstrated on continuum robots, our methodology is broadly applicable and holds the potential to advance complex autonomous systems across diverse fields such as aerospace, chemical engineering, and neuroscience.