New Machine-Learning Approach for More Effective Robot Control
Researchers from MIT and Stanford University have developed a new machine-learning method that could enhance the control of robots, such as drones and autonomous vehicles, in dynamic environments. This technique aims to enable robots to adapt to rapidly changing conditions, such as slippery roads or strong winds, by incorporating structure from control theory into the learning process.
According to Navid Azizan, an assistant professor at MIT, the focus of their work is to learn the intrinsic structure within the dynamics of a system. By jointly learning the system’s dynamics and unique control-oriented structures from data, they can create controllers that are more effective in real-world scenarios. Their approach allows for the immediate extraction of an effective controller from the learned model, resulting in better performance and faster learning in dynamic environments.
Unlike other machine-learning methods that require separate steps for learning a controller, the researchers’ approach integrates the controller into the model. Additionally, their technique requires fewer data for learning, making it more efficient. The researchers draw inspiration from how roboticists use physics to derive simpler models for robots, which in turn inform control logic.
The researchers’ technique has been tested and has shown promising results. The controller extracted from the learned model closely followed desired trajectories and outperformed other baseline methods. Furthermore, the method was data-efficient, achieving high performance with limited data.
The team believes that their approach can be applied to various dynamical systems and have plans to develop more physically interpretable models in the future. The integration of system dynamics, controllers, and control-oriented structure in a joint learning algorithm is seen as a significant contribution to the field of non-linear feedback control.
This research is supported by the NASA University Leadership Initiative and the Natural Sciences and Engineering Research Council of Canada. The findings will be presented at the International Conference on Machine Learning (ICML).