Title:
Fast Sensitivity Computations for Trajectory Optimization
Fast Sensitivity Computations for Trajectory Optimization
Author(s)
Arora, Nitin
Russell, Ryan P.
Vuduc, Richard W.
Russell, Ryan P.
Vuduc, Richard W.
Advisor(s)
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
Gradient based trajectory optimization relies on accurate sensitivity information to
robustly move a solution towards an optimum. Computational complexity of sensitivity
calculations increases exponentially for higher problem dimensions and
orders. Hence, the computation of these sensitivities is traditionally a major speed
bottleneck in trajectory optimization and targeting algorithms. We propose to use
Nvidia's GPU (Graphics Processing Unit) to rapidly calculate the derivatives in
a multilayer, parallel, and heterogeneous way while the CPU (Central Processing
Unit) sequentially computes the less expensive state equations. The proposed tool
computes both the first and second order analytic sensitivities on the GPU with
double precision accuracy. For an example trajectory propagation, we demonstrate
overlapped computations such that sensitivities are calculated almost for
free compared to the conventional CPU implementation.
Sponsor
Date Issued
2009-08
Extent
Resource Type
Text
Resource Subtype
Paper
Rights Statement
Unless otherwise noted, all materials are protected under U.S. Copyright Law and all rights are reserved