Continual learning framework lets soft robots adapt without retraining

Researchers Nilay Kushawaha, Muhammad Sunny Nazeer, Baljinder Singh Bal, Cecilia Laschi, and Egidio Falotico have introduced a continual learning framework for controlling modular soft robots (MSRs). Unlike existing methods that require controllers to be retrained from scratch whenever the robot's morphology changes, this framework incrementally adapts while preserving previously acquired knowledge. The approach is validated through closed-loop trajectory tracking experiments on a simulated tendon-driven soft robot and a real-world three-module pneumatic soft robotic arm. Additionally, for fixed-configuration MSRs, the framework can be deployed in a distributed manner to learn module-specific dynamics, enabling localized control and improved precision. A reaching experiment demonstrates the controller's ability to selectively activate only necessary modules to reach a virtual target, reducing computational overhead. The work addresses key challenges in MSR control: nonlinear dynamics, modeling complexity, and hyper-redundant structure. Published on arXiv (2607.06740), the research sits at the intersection of robotics (cs.RO) and artificial intelligence (cs.AI).