Evidence and limits
Part of the Aurora NERVE overview: learned motion prediction for the Aurora spaceplane.
This project treats “what we can show” and “what we hope to show” as different
categories and keeps them apart.
Demonstrated in software
- Versioned sensor, observation, replay, step-audit and flight-log contracts.
- Causal forecast recording: predictions are stored before their outcomes.
- Independent offline trajectory scoring, with separate alignment and scoring
intervals, demonstrated on complete synthetic logs.
- Refusal when a complete state is not observable from payload-local sensing.
- Independent, fail-closed assurance gates (confidence, ensemble disagreement,
physics, finite-value, command and rate).
- Local monotonic scheduling, with network time kept out of scheduling decisions.
- Append-only, hash-chained recording of raw packets and per-tick decisions.
- Three interchangeable simulation plants behind one contract: an analytical
surrogate, a 6-DoF JSBSim X-15, and a provenance-gated CFD-table plant.
- A test suite of over 1,200 tests, run with lint, type-checking and an
environment-contract check before every change.
Measured, with honest results
- Simulator gyro-history study: learned linear forecasts beat naive
persistence on the X-15 surrogate. A benefit over strong filtered baselines is
not established, and the project does not claim one.
- A direct forecasting benchmark was recorded as failing its fixed criterion.
The failure was kept, not tuned away.
- The learned control policy has not yet beaten the zero-residual baseline
(the scheduled controller acting alone). The project reports this openly.
- An evaluation-scenario defect was found that had made controller outcomes
independent of the controller. It was fixed, and earlier results were declared
non-comparable rather than reused.
Not yet established
- Physical sensor capture under a qualified acquisition path.
- The trained flight predictor and calibrated uncertainty intervals.
- Flight hardware integration and environmental qualification.
- Any Aurora-representative aerodynamic or flight-dynamics model.
- Any flight benefit, pilot comparison, or performance above Mach 3.7.
Why this matters
The value proposition for flight-test and instrumentation teams is evidence they
can trust when deciding whether a learned predictor deserves integration. That
includes the conditions where simpler methods do better.