Can my model beat RAPTOR?

The final comparison includes the possibility that a classical controller wins.

Read the companion now. The recording is forthcoming; “Free” identifies its planned release track.

RAPTOR is an external adaptive quadrotor policy with published research results. It is not automatically a compatible binary for the course’s Bolt build, and it outputs motor commands from its own onboard controller, so it flies as a separate onboard deployment rather than through the shared off-board route. Validate checkpoint, action semantics and deployment support before including it in a flight campaign; otherwise keep an explicit unavailable entry with the reason.

Compare supported controllers on nominal, payload, fan and safe low-battery conditions with randomized order and repeated flights. Report error, failures, recovery and command effort from the same harness. Lighthouse comparisons belong in the same calibrated volume, with positioning limitations acknowledged.

The final artifact is a reproducible table and synchronized evidence, whatever the ranking. A PID win can reveal a narrow task, effective tuning or an unnecessary learned model. An adaptive win becomes more credible when the classical baseline was well tuned and the missing or failed trials remain visible.

Take this into your next build

The course succeeds when the comparison is reproducible, not when a particular model wins.

Put the idea to work

Fly the trained model

Bring every controller through the same off-board interface, verify a compact onboard export, and publish a hardware comparison even when the classical controller wins.

Build module 09