Thank you for visiting nature.com. For the best user experience on our site, we recommend using an up-to-date browser. In this paper, a novel machine learning framework is developed to predict the success of vertical-wall perching for flying robots with spines. Traditional methods lack efficiency in analyzing perching dynamics, which inspired the development of this new approach. This framework optimizes robot design parameters and control strategies, ensuring stable perching while reducing design time and costs.
Perching is an essential capability of animals to land and stay on various surfaces. Animals exhibit diverse perching techniques, such as flies landing upside down on ceilings, geckos crashing into tree trunks, and bats landing on cave ceilings. Robots, especially aircraft, face challenges in perching on vertical surfaces. Developing perching mechanisms inspired by animals can enhance robot mobility, robustness, and working time, enabling applications in extreme environments like search-and-rescue operations.
Several studies have explored novel perching mechanisms for robots, but predictive frameworks for perching outcomes are lacking. Wall-perching is a critical challenge for flying robots due to rapid maneuvers and strict velocity constraints. Machine learning implementations, such as deep reinforcement learning, have shown promise in enhancing autonomous capabilities in flight control domains.
This work presents a machine learning-based predictive framework for vertical-wall perching and flying robot design. The framework combines knowledge-based modeling, computational impact dynamics, and experimental data. By training mixed datasets comprising simulation and experimental data, a data-driven model is established to predict perching outcomes. This model guides robot design and prevents landing failures, ultimately enhancing perching capabilities.
The study focuses on designing a spiny flying and wall-climbing robot with three subsystems: flight, landing, and attachment systems. The landing process involves four stages, from approaching the wall to successfully perching. By implementing the machine learning model, the robot can predict landing outcomes in real-time and adjust its flight states accordingly for successful perching.
Experiments were conducted to collect data on successful and failed landing outcomes, which were used to train and validate the machine learning model. Moreover, a closed-loop optimization process was employed to fine-tune the robot’s landing rod structure for optimal performance.
Overall, the machine learning-based framework presented in this study offers a more efficient and accurate way to predict landing outcomes for flying robots. This approach not only improves perching capabilities but also guides the design and optimization of future robot structures for enhanced performance in complex environments.
About the Author
