Research/ SPACE AI & AUTONOMY
astrograd a modular, differentiable astrodynamics library for massively parallel simulations and code generation
Valentin Preda, Nazar Misyats, Frederik Markus, Lorenzo Capra
Guidance, Dynamics & ControlOpenAlex keyword: guidance, navigation and control
We introduce astrograd, a Python astrodynamics library that combines automatic differenti-ation, hardware-accelerated parallel execution and embedded code generation with modelsspecific to Guidance, Navigation and Control applications. The library is built around asingle idea: assemble a differentiable simulator once from reusable modules, then run itunmodified across many numerical computing backends. Simulators built with astrogradcan be coupled with deep-learning components and executed with massive parallelismon machine-learning frameworks such as JAX, PyTorch or TensorFlow. The same modelswork directly with CasADi for optimal control and estimation, or can be compiled to em-bedded code for flight processors. astrograd ships with validated modules for the orbitalenvironment, rigid body dynamics, ephemerides, frame and time conversions, numericalintegration and finite-state machines needed to build sophisticated spacecraft simulators forcontrol system development and verification. astrograd is released under a permissive open-source license for the aerospace community, together with representative examples andtutorials covering hardware accelerated robust satellite control design, frequency-domainanalysis, re-entry corridor analysis, uncertainty propagation, trajectory optimization andembedded code generation.