Learning to control a soft robotic manipulator under uncertainty and unforeseen changes in robot–environment interaction

Safe and efficient robot–environment interaction is a critical yet challenging problem, particularly in the presence of uncertainty and unforeseen changes. Soft robotics, known for its inherent compliance, enables safe interactions with the environment but often suffers from inefficient control performance. Meta-learning, which leverages a robot’s prior experiences in different environments, offers the potential for rapid adaptation to similar conditions with minimal observations. Building on this principle, this work develops a learning-based control approach to enhance the control efficiency of soft robotics. Specifically, a flexible meta-learning model structure is designed to address robot–environment interactions across different situations. An uncertainty-aware optimal control policy, equipped with stability guarantees, is carefully crafted to achieve desired performance. The proposed approach is validated on two soft robotic systems: a pneumatic cable-driven soft manipulator and a rod-driven soft robot. Experimental results demonstrate that these robots can rapidly adapt to varying environmental situations and effectively achieve control objectives, even in the presence of random and unforeseen disturbances. Furthermore, comparisons with other learning-based and physics-based control methods highlight the superiority of our approach in terms of faster adaptation, improved stability, and higher accuracy. This work provides a feasible control approach for soft robots to handle uncertainty and adapt to unforeseen changes in robot–environment interactions.