Research
My research lies at the intersection of estimation theory, convex optimization, and robust control, which are essential ingredients in the navigation of autonomous robots. Before a robot can plan or act, it must know where it is, using measurements that are noisy. In real world, the measurement noise is often non-Gaussian. Under such noise the standard Kalman filter can perform poorly, and the many robust estimators proposed as alternatives rarely come with guarantees.
A central theme of my current work is the design of certifiably stable state estimators (e.g., Kalman-type filters) for systems under non-Gaussian noise. I use tools from robust control and convex optimization to address two questions:
- How can we design new estimators that outperform the standard Kalman filter and existing non-Gaussian estimators?
- How can we certify the stability of non-Gaussian estimators, i.e., guarantee that their estimation error remains bounded?
I am particularly interested in the second question, since stability guarantees for most existing non-Gaussian estimators are limited or rely on restrictive assumptions. Previously, I also worked on vision-based navigation and analytical dynamics. Brief summaries are provided below.
-
State estimation under non-Gaussian noise
The Kalman filter is BLUE: best linear unbiased estimator for both Gaussian and non-Gaussian noise distributions. However, under non-Gaussian noise its performance degrades significantly, thus we need robust nonlinear estimators. We design estimators that stay accurate under such noise at a computational cost close to that of the Kalman filter, and I prove that their estimation error remains bounded. [paper1] [paper2] [paper3]
-
Analytical dynamics
Guidance, navigation and control of a dynamical system all start from its equations of motion. We developed an efficient form of Volterra’s equations for multibody systems that gives the minimum number of first-order equations needed to describe their motion. [paper]