Researchers from MIT and Stanford University have developed a new machine-learning approach that could improve the performance of robots, such as drones and autonomous vehicles, in rapidly changing environments. This approach incorporates control theory into the learning process, allowing for more effective control of complex dynamics caused by factors like wind impacting a flying vehicle’s trajectory.
The researchers aim to learn the intrinsic structure within a system’s dynamics, enabling the design of more effective controllers. By learning this structure and the system’s dynamics simultaneously from data, the researchers can create controllers that perform better in real-world scenarios.
Unlike other machine-learning methods, this approach immediately extracts an effective controller from the learned model, eliminating the need for separate controller derivation or learning. It also requires less data to learn an effective controller, resulting in faster performance in dynamic environments.
The researchers draw inspiration from how roboticists use physics to derive simpler models for robots. They seek to identify useful structure from data that guides the implementation of control logic.
Learning a controller for a robot is challenging, especially when the system is too complex to be manually modeled. Manual modeling captures the system’s structure based on physics, but complex factors like aerodynamic effects are difficult to derive manually. Machine learning fits a model to data, but these models often lack control-based structure necessary for effective control.
Most existing approaches learn a separate controller for the system from data. However, the researchers’ approach integrates the dynamics model with a prescribed structure that supports control. They extract a controller directly from the dynamics model, resulting in improved trajectory tracking.
The researchers’ technique demonstrated high data efficiency and outperformed baseline methods in tracking capability. It achieved effective modeling of dynamic vehicles using fewer data points compared to other methods. This efficiency makes the technique valuable in rapidly changing environments.
The approach is generalizable and applicable to various dynamical systems, including robotic arms and free-flying spacecraft in low-gravity environments.
Future research aims to develop models that are more physically interpretable, allowing for the identification of specific information about dynamical systems. This could lead to even better-performing controllers.
The research is supported by the NASA University Leadership Initiative and the Natural Sciences and Engineering Research Council of Canada.