Abstract:
Many problems in robotics can be formulated as quadratic programs (QPs). In particular, model-predictive control problems often involve repeatedly solving QPs at very high rates (up to kilohertz). However, while other areas of robotics like machine learning have achieved high performance by taking advantage of parallelism on modern computing hardware, state-of-the-art algorithms for solving QPs are still inherently serial. In this talk, I’ll show how we formulated the Alternating Direction Method of Multipliers (ADMM) algorithm for solving QPs such that it maps mathematically to the same operations as a neural network with rectified-linear unit (ReLU) activations. I’ll compare this new algorithm to state-of-the-art solvers such as OSQP and ProxQP on high-dimensional whole-body MPC problems and show where it excels, and what its limitations are. Finally, I’ll discuss how this work is a first step toward finding a balance between offline and online compute to better leverage parallel hardware architectures.
Committee:
Zachary Manchester (advisor)
Aaron Johnson
Lorenz Biegler
Kevin Tracy