We aim to understand how biological movements are controlled, and synthesize control systems that achieve similar performance. Our approach is based on optimal control. Biological movements can be modelled in detail using optimality principles - which is not surprising given that motor behavior is shaped by iterative optimization processes such as evolution, learning, adaptation. Similarly, the best way to engineer a complex control system is to specify a high-level performance criterion and leave the details to numerical optimization. In both areas, the main difficulty lies in actually performing the optimization. Thus our focus is on developing more powerful methods for optimal control, and applying them to harder problems.
Evangelos Theodorou, former postdoc in the lab, is now Assistant Professor of Aerospace Engineering at Georgia Tech. Congratulations! We hope he enjoys grant writing and teaching and meetings :)
Convex structured controller design
Dj Krishnamurthy found a way to optimize stability under arbitrary convex constraints on feedback gains, including sparsity and bounds. The resulting optimization problem is convex, yet it approximates the H2 and Hinf criteria that are known to be NP-hard.
Analytically-invertible dynamics with contacts and constraints
Emo Todorov realized that his earlier convex contact model can be inverted analytically. As a result, the inverse dynamics of a humanoid with multiple active contacts can now be evaluated around 1000 times faster than real-time.