Trajectory Optimization and Morphology Control for Soft Growing Robots via Sim and Real Co-Training

As a novel class of bioinspired soft robots, soft growing robots can navigate by interacting with their surroundings and exhibit strong adaptability in unstructured and complex environments owing to their high compliance. However, existing soft growing robots exhibit limited morphology control, and current motion planning methods do not fully exploit morphology-environment interactions, leading to inaccurate trajectory planning. Moreover, learning-based planners rely heavily on simulation-generated data, while the nonlinearity of soft growing robots introduces significant sim-real discrepancies. To address these challenges, we propose a sim-real co-training framework that integrates morphology control and trajectory optimization in soft growing robots, using large-scale simulated and limited real data to enhance real-world performance. The algorithm adopts a hierarchical learning framework: the first-stage network employs an imitation learning network for pre-movement trajectory prediction and optimization. The second-stage network performs error correction on the trajectories predicted by the first-stage network. It integrates a tendon-driven mechanism and employs offline reinforcement learning to achieve morphology control. Experimental results show that the proposed strategy significantly improves planning accuracy and generalization. It further enhances the robot's motion accuracy, decision-making, and error-correction capabilities. Overall, it shows strong potential to enable new forms of embodied operation in soft growing robots through morphological intelligence. The source code and partial datasets are publicly available at https://github.com/wuhaoran6/soft-growing-robot-motion-planning Note to Practitioners - For soft growing robots operating in confined environments, this paper proposes a path optimization and morphology control method and develops a sim-real co-training framework that integrates limited real-world data with large-scale simulated data, thereby improving planning accuracy and generalization performance in real-world scenarios. In addition, the algorithm adopts a hierarchical learning framework that decouples trajectory prediction from morphology control. The proposed algorithm is well-suited for environment-interaction-based motion planning of soft growing robots in confined environments, and it effectively exploits morphological adaptation and environmental constraints to guide motion optimization, thereby enhancing motion planning accuracy, autonomous decision-making, and error correction capability. This work provides a new pathway toward enabling embodied operation in soft growing robots.