Condensed PIPG Sequential Convex Optimization for Reusable-Rocket Powered Landing with Strong Aerodynamics

What the math gives to ML

The transferable contribution is an exact reduced-space treatment of large equality-constrained discretized dynamics, combined with a projected primal-dual gradient kernel whose cost is dominated by fixed-size matrix-vector products rather than a generic conic solve. This structure can be applied to neural ODE, SSM, or differentiable-controller training when hidden-state collocation constraints, terminal conditions, or safety constraints create many redundant state variables. The most promising adaptation is to linearize the neural dynamics at each outer Gauss-Newton or sequential-convex-programming iteration, eliminate intermediate state increments by forward sensitivity recursion, and run a projected primal-dual method only on parameters, controls, and the small terminal-constraint space. Its value should be tested specifically on constrained neural dynamical systems rather than unconstrained Adam workloads.