Rapid Exploration of the Design Space During Automatic Generation of Kalman Filter Code

State estimation is a core capability for autonomous systems such as satellites and planetary rovers. Kalman filters provide a computationally efficient way to determine the values of state variables (e.g. position, velocity) from noisy measurements. The Automated Software Engineering group at NASA Ames Research Center has previously developed a number of systems for generating program code in NASA-relevant domains, including AMPHION/NAIF (Whittle et al., 2001) for generating code for mission planning and AUTOBAYES (Fischer and Schumann, 2003) for generating data analysis code. In this paper, we outline a program generation system, AUTOFILTER (Whittle and Schumann, 2005), which has been developed at NASA Ames Research Center. AUTOFILTER synthesizes (i.e. generates) Kalman filter code. It takes as input a textual specification - a description of the mathematical model underlying the Kalman filter - and automatically generates code suitable for compilation using MATLAB libraries (for prototyping and testing the filter) or standalone C code (for deploying the filter). We describe how AUTOFILTER assists the iterative development of Kalman filters in various ways: permitting changes in the mathematical model underlying the filter to be rapidly realized as code and tested, different Kalman filters to be synthesized from the same model, code automatically assessed for computational performance, and approximating assumptions applied to the code in order to improve efficiency. AUTOFILTER provides assistance for ensuring that the generated code is correct by generating program documentation and correctness certificates in addition to the code itself