I designed and evaluated a hybrid landing controller for a simulated rocket subject to actuator degradation, mass uncertainty, and off-nominal approach states. The design combines constrained finite-model MPC, local LQR feedback, recovery supervision, and dedicated touchdown logic.
- A shared-input MPC plans against a finite set of actuator-effectiveness models while enforcing local state and actuator constraints.
- Online effectiveness estimates adjust hover trim and available control authority; clipped LQR feedback supplies a small local correction.
- A hybrid supervisor handles recovery outside the local model region and switches to braking, touchdown gating, and contact settling near the deck.
All three withheld malfunction tests ended in two-leg landings. Among 124 full-credit submissions, this controller had the fastest end-to-end evaluation (191 submissions overall).
End-to-end evaluation refers to execution of the complete evaluation notebook, not an individual MPC solve. The recorded trajectories below use simulator output; the shaded area is the deck and the markers are final positions.
The executable supervisor reference captures mode selection, entry hysteresis, solver fallback, touchdown commitment, and contact-settle handoff behind generic policy interfaces. The control formulation documents the sanitized MPC and LQR structure.
python3 -m unittest discover -s tests -vControl: robust MPC, LQR, hybrid supervision, online effectiveness estimation, and constrained touchdown control.
Tools: Python, CVXPY, OSQP, simulation-based validation, and unit testing.
This simulator study is not a flight-certified system. Exact models, constants, test cases, simulator code, and submission code are intentionally omitted; see the disclosure boundary.