Data-Efficient Learning Control of Continuum Robots in Constrained Environments

This research investigates learning-based control of continuum robots in constrained environments without relying on analytical models. We propose a data-efficient stochastic control strategy incorporating online model updates to achieve precise manipulation even when arbitrary robot deformations occur due to environmental interactions. A localized Gaussian process regression approach accounting for state stochasticity is first presented to approximate the forward kinematics. The learned model enables uncertainty-aware stochastic predictions via the proposed scaled unscented transform (SUT)-based method for efficient exploration. Leveraging new data, online model updates are performed in a highly sample-efficient manner. Furthermore, a probabilistic model predictive control approach integrating the learned models and chance constraints based on Chebyshev's inequality is developed for searching an optimal control sequence. Simulations and experiments are performed to demonstrate the effectiveness of the proposed approach for controlling continuum robots in constrained environments using limited observational data. Note to Practitioners —The motivation of this research is to solve the problem of controlling continuum robots in constraint environment. The flexibility of continuum robots significantly affects the manipulation accuracy, and the interaction between the continuum robot and environmental constraints can also lead to unpredictable behavior. Learning control methods that rely only on sensory data, provide a feasible solution to the aforementioned problem. However, current methods lack sample efficiency and the capability to handle unknown environmental constraints. This research proposes a learning control method which can control a flexible continuum robot in constrained environments with high data-efficiency and robustness even when the robot shape undergoes sudden deformations due to contact with obstacles.