reylink

Recorded run · 2026-10-17 · simulation

Claude planned a ten-drone survey. reylink refused the unsafe plan.

Claude, an AI model, wrote two survey plans for ten PX4 drones, the first made unsafe on purpose. reylink refused it at preflight, before any motor ran, then proved the second plan and flew it. This is that run, filmed.

mission-film-20261017 · gazebo 3d + atlas 44 s · sped up 1–4× · no sound

Simulation, filmed live. Captured from Gazebo's 3D view (PX4 SITL) and from Atlas, reylink's operator UI, connected read-only. Playback speed and the mission clock are on screen.

The run in four figures

10drones flew the AI-written survey and landed on their own padsPX4 v1.17 SITL
1unsafe plan refused at preflight: the paths of px4_4 and px4_6 conflictedno motor ran
29runtime holds when the guard could not prove 2 m separation; every one clearedin flight
3.84 mclosest pair across the whole run; the enforced floor is 2.0 mGazebo world truth

What the film shows

  1. T+0 sTwo AI-written survey plans go to preflight. Plan 1 is refused at once: the paths of px4_4 and px4_6 conflict.
  2. T+3 sPlan 2 passes preflight with separation proven, and the launch starts.
  3. T+8 sTakeoff: all ten leave their pads and climb before any route starts.
  4. T+19 sRoutes start: each drone flies to its own survey column.
  5. T+21–45 sAtlas, connected as a read-only observer, tracks all ten live while the mission runs through the operator API.
  6. T+45–68 sSurvey lanes 12 to 20 m above takeoff. Routes complete one by one, 10 of 10 at T+68 s.
  7. T+72 sOperator LAND: each drone descends onto its own pad.
  8. T+97 sComplete: all ten back on their pads, disarmed, confirmed.

What reylink checked

Before launch. Preflight has to prove separation for every launch column, approach and route before a drone arms. Plan 1 failed that proof on one pair of paths, px4_4 and px4_6, so it was refused and the reason was recorded. Plan 2 passed with separation proven, along with the armability and full launch-geometry checks, and the launch started.

In flight. Every setpoint is checked against a 2 m separation floor, with predictive refusal. The runtime guard paused drones 29 times when it could not prove they would stay 2 m apart, and every pause cleared. The closest pair in the run stayed 3.84 m apart.

On the record. Every position in the replay is from the Gazebo world-truth log, and the film is a screen capture of the same run. Every command, decision and refusal is logged with its reason.

The AI, bounded. reylink ships no AI model of its own. It is the envelope any model runs inside: AI proposals enter through the operator layer and cross the same gate as every other command, so a plan from Claude is checked exactly like a plan from a person.

Test conditions: PX4 v1.17.0 software-in-the-loop with Gazebo Harmonic, ten simulated drones, run claude-split-survey-n10-20261017T1355Z (2026-10-17, 13:55 UTC). Positions resampled to 2 Hz from Gazebo world truth. Nothing on this page was flown on real aircraft.