Part of the Aurora NERVE overview: learned motion prediction for the Aurora spaceplane.
Measure how far ahead a learned model can predict measured rotational motion during high-speed flight, and whether its uncertainty identifies when its forecasts fail. The intended output is a reproducible report of error, availability, uncertainty and remaining lookahead across the conditions actually flown. Positive, negative and inconclusive outcomes are all reportable.
The target is the payload’s own later three-axis gyro reading. This measures prediction of sensor output. Sensor bias, vibration and filtering effects are kept in the result, not removed. Claims about vehicle body rate would also need a verified mounting transformation.
The 0.375 s primary horizon lies in the middle of the proposed range. It is a research operating point, not a measured latency requirement of any user.
| Question | Method |
|---|---|
| Does learning improve prediction? | Compare a compact GRU correction model and a ridge history model with raw persistence, filtered constant rate and causal trend. Choose the decisive conventional comparator on development data. Primary endpoint: paired per-flight gyro vector RMSE difference at 0.375 s on common evaluable origins, before confidence selection. Report per-axis and tail errors and missing forecasts. |
| How much lookahead is actually available? | Evaluate all horizons after measured sensing, filtering and issue delay. Target 90% issuance among sensor-valid, warmed-up origins, and also report availability over every scheduled origin. Compare methods at matched issuance. |
| What accuracy does it support? | Report the longest horizon that meets exploratory per-axis 95th-percentile error bands of 0.1, 0.5 and 1.0 deg/s. These bands are for sensitivity analysis only. Whether a result is useful depends on a user’s own tolerance. |
| Is uncertainty informative? | Ensemble or quantile intervals calibrated on separate sessions. Measure coverage, width, confidently wrong cases and refusals against a conventional selector that uses recent causal errors. |
Known motion structure enters through causal rate extrapolation and rotation kinematics. The learned correction is not presented as an identified aerodynamic force or turbulence estimate.
2–3 boost-glide flights with an identical frozen configuration. Ground commissioning comes first. The first flight evaluates the integrated system, the second measures repeatability, and a third adds exposure if allocated. No special manoeuvre or fixed supersonic duration is required.