Detector in, frame in,
command out.

§ 01Single-file example

from vq1_completion_pilot import (
    ApexDetectorChain, CompletionPilot, CameraIntrinsics, Telemetry
)
from pathlib import Path

intr = CameraIntrinsics(width=640, height=360, fx=320, fy=320, cx=320, cy=180,
                        tilt_deg=20.0, gate_width_m=1.5, gate_height_m=1.5)
detector = ApexDetectorChain(
    detector_path=Path("models/apex_detector_best.onnx"),
    keypoints_path=Path("models/apex_keypoints_best.onnx"),
)
pilot = CompletionPilot(intr=intr, detector=detector)

# ------- one frame, one command -------
frame_bgr = cv2.imread("test_frame.png")
tele = Telemetry(
    attitude_q=(1.0, 0.0, 0.0, 0.0),
    body_rates=(0.0, 0.0, 0.0),
    body_accel=(0.0, 0.0, 9.81),
    t_sim=0.0,
)
cmd = pilot.step(frame_bgr, tele)
print(cmd)   # ControlCmd(throttle=0.55, roll=0.0, pitch=0.01, yaw=-0.04)

§ 02What each line does

• CameraIntrinsics — camera model. Replace with real AIGP-sim intrinsics when credentials drop.
• ApexDetectorChain — YOLO11n detector + YOLO11n-pose keypoints, loaded once, reused per frame. Stub returns None until real weights wired.
• CompletionPilot — deterministic PID over gate pose. Zero learning. Ships as-is for VQ1.
• pilot.step — full per-frame pipeline. Returns a ControlCmd for Throttle/Roll/Pitch/Yaw.

§ 03Swapping in the real sim

Replace StubSimClient in vq1_completion_pilot.py with the real AIGP sim adapter when credentials are released:

class AIGPSimClient:
    def connect(self) -> None: ...
    def next_frame(self) -> tuple[np.ndarray, Telemetry]: ...
    def send(self, cmd: ControlCmd) -> None: ...
    def close(self) -> None: ...

The rest of the pipeline is transport-agnostic.

§ 04Troubleshooting