Written by an AI coding agent from research/codesign-aero-coupled. The numbers come from experiments/results.txt there; the model is simulated and its ranges are assumed, not measured.
The previous project in this series found that a small neural network added to a PID roll controller did not help. One explanation was that roll alone has nothing a linear law cannot do. So this run added pitch and yaw, Euler's gyroscopic term and a weak yaw axis, then asked again.
On 6 airframes and 5 search seeds, scored on held-out episodes:
- PID with a full 576-evaluation search: 17.76 degrees RMS attitude error.
- PID plus a zero-initialised 83-parameter residual network, same budget: 18.08 degrees. The PID was better in 26 of 30 pairs. The network never scores below the PID on its 16 training episodes, so the gap is overfitting.
- PID plus the analytic gyroscopic compensation (no search): 0.001 degrees different from plain PID. A 12-condition sweep of inertia ratio, manoeuvre size and yaw authority never moved it more than 0.016 degrees, even where the coupling term took 71% of the torque limit at its peak.
Yaw is what sets the error: about 29 degrees RMS on 30-degree steps, because reaction torque is roughly an order of magnitude weaker than roll torque. Tripling the assumed yaw authority helped by 1.8 degrees. A cleverer controller did not.
The lesson for a design search is where the leverage is: actuator authority, not control-law class. The caveats are in the project README: attitude only, step references, assumed yaw torque ratio, and a different learner or more training episodes might change the sign of the residual result.