# Control discovery

> An open research program using autonomous AI research agents to discover, falsify, and refine control strategies for uncertain nonlinear systems.

Canonical page: https://aerospace-superintelligence.com/control/

Status: founding phase. Milestones are intentions, not achieved results.

## Mission

Build a reproducible discovery engine for control laws that are more robust, efficient, or capable than established baselines.

Research question: can AI discover a controller that outperforms a strong, fairly tuned baseline under unfamiliar conditions, and reproduce the result?

AI-generated engineering ideas are treated as hypotheses, not accomplishments. No performance claim is published without a traceable experiment.

## Research agenda

1. **Nonlinear control discovery**: interpretable adaptive, robust, and model-based strategies beyond textbook examples.
2. **Automated falsification**: randomized environments, adversarial disturbances, and carefully separated test sets.
3. **Efficient implementation**: performance, inference cost, and latency, with eventual embedded deployment in mind.

Research cycle: hypothesize → simulate → challenge on held-out conditions → promote only improvements backed by repeatable evidence.

For aircraft problems, the plan is to challenge candidates with the [flight tools](https://aerospace-superintelligence.com/tools.md): nonlinear simulation with turbulence, Monte Carlo dispersions, and MIL-F-8785C flying-qualities criteria, run independently of the agents that proposed them.

## Roadmap

- I. Foundation (current focus): deterministic plant models and baseline controller comparisons.
- II. Search (planned): autonomous candidate generation with independent validation.
- III. Robustness (future): cross-system benchmarks and rigorous holdout testing.
- IV. Hardware (long-term): deploy verified controllers on dedicated compute hardware.

Sister program: [Silicon](https://aerospace-superintelligence.com/silicon.md).