Imagine a bustling city where unmanned aerial vehicles (UAVs) seamlessly navigate through the urban jungle, autonomously avoiding skyscrapers, responding to sudden gusts of wind, and effortlessly adapting their paths to ensure swift deliveries. This vision of the future hinges on advanced control systems that can learn and adapt to new tasks without human intervention. At the heart of this innovation lies the groundbreaking research by EE Professor James Anderson and his student researchers at Columbia University, who have recently been honored with the Best Paper Award at this year’s Learning for Dynamics and Control Workshop (L4DC).
The awarded paper (accessible here), delves into the complexities of "meta-learning" or "learning to learn," a field that has garnered significant attention and success in recent years. Professor Anderson’s research provides precise conditions under which meta-learning approaches can succeed and characterizes the computational effort required for success. By adapting the well-known Model-Agnostic Meta-Learning (MAML) algorithm to dynamical systems and control engineering, the team demonstrated that they could produce a controller with performance guarantees close to task-optimality, factoring in task heterogeneity.
“Meta learning is a large and diverse field and has celebrated many success stories over the last few years,” says Professor Anderson. “However, the majority of the results tend to be empirical in nature. Our paper was able to precisely determine when such an approach can succeed and characterize the amount of computation required to succeed.”
Professor Anderson emphasizes that this research is not specific to UAV's, as the theory is very general. It can be applied to any dynamical system that has (approximately) linear dynamics, and all the results go through seamlessly.
“This is in stark contrast to the more experimental work in this area, which often requires extensive hand-tuning and even then you don't know if meta-learning will work,” Professor Anderson says.
The team’s work stands out not only for its theoretical contributions but also for its potential practical applications. Their algorithm shows promise in stabilizing systems across all tasks at each iteration of the design process and provides an upper-bound on the rate at which it converges to the optimal controller.
L4DC, hosted this year at the University of Oxford, aims to create a new community that rigorously integrates machine learning, control theory, and optimization. As the explosion of real-time data from physical world devices continues, rethinking the foundations of these disciplines becomes crucial. The workshop's mission aligns perfectly with the goals of Professor Anderson and his team.
Reflecting on the award, Anderson emphasized, “We’re thrilled about this project and are eager to implement it on real robotic systems, a collaboration we’re currently pursuing with Professor Brian Plancher (EE affiliated faculty) at Barnard.”
Read the full paper here.
