Skip to main content
Read this before the first flight of any aircraft running ADOS. The short version: ADOS does not fly your aircraft. The flight controller does. ADOS gives you a link to it, a checklist against live telemetry, and panels to configure it, but every protective behaviour that matters when something goes wrong lives in the FC and is off until you set it.

What ADOS does not do

The agent has no flight envelope of its own. There is no companion-side geofence and no companion-side return-to-launch. If the FC is not configured to protect the aircraft, nothing else will.
  • No geofence enforcement in the agent. The agent can republish a geofence breach the FC reports, as a notification. It never commands anything in response.
  • No autonomous RTL. rtl exists only as a command you send.
  • No pre-arm checks of its own. The agent relays your arm request to the flight controller and reports back what the FC decided, including the FC’s own reason text when it refuses. Every arming rule lives in the flight controller. If the FC’s pre-arm checks are disabled, nothing in ADOS re-adds them.
  • The swarm setpoint loop stops rather than substitutes. When the vehicle is disarmed, the FC has left GUIDED, or the neighbour table has gone stale, it stops emitting and the FC holds on its own terms.

Before the first flight

1. Set the failsafes on the flight controller

Open the drone’s Failsafe panel and the Power panel in Mission Control. At minimum, decide the answer to each of these before you fly: Two of these are easy to miss. The critical battery tier is on the Power panel, not the Failsafe panel, so setting only the low tier from the Failsafe panel leaves the second stage unset. And no panel edits RTL_ALT, so set it in the generic parameters editor. A default RTL_ALT that is lower than the trees between the aircraft and home is the classic way to lose one. On a plane or VTOL the RC failsafe parameters are THR_FAILSAFE and THR_FS_VALUE instead. On PX4, the panel writes the equivalent PX4 parameters where a mapping exists: battery to COM_LOW_BAT_ACT and BAT_LOW_THR, the fence to GF_ACTION, GF_MAX_VER_DIST, and GF_MAX_HOR_DIST, and RTL_ALT to RTL_RETURN_ALT. The panel does not surface PX4’s link-loss actions, so set those in the parameters editor. On Betaflight the equivalents are BF_FS_DELAY, BF_FS_OFF_DELAY, BF_FS_THROTTLE, and BF_FS_PROCEDURE. On iNav, the failsafe panel exposes failsafe_nav_mode, failsafe_min_distance_behaviour, and failsafe_min_distance.

2. Understand what GCS failsafe will and will not do for you

This one is specific to running a companion computer, and it surprises people. The agent sends a heartbeat to the flight controller once a second, identifying itself as an onboard controller. That is deliberate: without it, ArduPilot would decide the GCS had vanished a few seconds after boot and trigger its GCS failsafe while the aircraft sat on the bench. The consequence is that the FC’s GCS failsafe is satisfied by the companion, not by your link to the aircraft. Losing video and telemetry to your laptop does not trip it, because the agent is still talking to the FC from a few centimetres away. FS_GCS_ENABL only fires if the agent itself dies or the serial link between the companion and the FC breaks. So do not treat GCS failsafe as your downlink-loss protection. RC failsafe, battery failsafe, and the geofence are the backstops that actually fire when you lose the aircraft. Configure those.

3. Work the pre-flight checklist

Mission Control’s cockpit checklist has 18 items, most of them verified against live telemetry rather than against your memory:
  • Hardware: battery above 20 percent, at least 3.7 V per cell, props secured, frame intact, motors free to spin
  • Software: 3D GPS fix, at least 8 satellites, EKF variances within limits, sensors healthy, firmware pre-arm pass
  • Environment: wind acceptable, airspace clear, launch area clear, observers briefed
  • Mission: flight plan loaded, geofence set, return-to-launch set, emergency procedures reviewed
Items with a telemetry source verify themselves, and drop back to pending when that telemetry goes stale. The rest you confirm by hand. A checklist belongs to one aircraft: completing it on one drone never vouches for another.
The checklist’s pre-arm item is derived from the sensor-health report, not from the flight controller’s own pre-arm text. Read the Health panel for the FC’s actual pre-arm verdict and its named blockers. The two can disagree, and the FC is the one that decides whether the motors spin.
There is a second, shorter checklist in the mission planner. That one is advisory, is not persisted, and does not gate arming. The cockpit checklist is the one that matters.

Arming

Arming is a skill in the cockpit Skill Bar, not a button that can be brushed. It requires the aircraft to be disarmed, and each flight-critical skill opens a confirmation sheet with its own gesture:
  • Arm is a slide-to-confirm. If the pre-flight checklist is incomplete, the sheet lists the open items and the slide stays inert until you turn on Override checklist. The override is recorded as a safety event. Skipping the checklist stays possible, because a field sometimes demands it, but it cannot happen quietly.
  • Disarm is press-and-hold.
  • Kill opens with its guard armed for 3 seconds. Only a hold started inside that window cuts the motors; if the window lapses, the request is cancelled.
Every gesture can also be completed by holding Enter or the bound gamepad button. Default hotkeys and the full Skill Bar are on Cockpit and Skills.

What a command confirmation means

A command is not reported as successful until the flight controller acknowledges it. The agent sends the command, waits about 1.2 seconds, and resends up to two more times before giving up. If the FC replies that it has accepted the command but needs time, the agent waits up to 8 seconds more for the outcome. If no acknowledgement arrives inside that budget, the agent reports the command as not observed. It never reports a success it did not see. Treat an unconfirmed command as one that may or may not have reached the aircraft, and check the vehicle’s actual state before sending it again. When the FC refuses a command, its own reason text is attached to the refusal. That text is the authoritative explanation, not anything ADOS infers.

The kill switch

The cockpit offers a kill switch alongside arm and disarm. It is not a stronger disarm.
The kill switch is an emergency motor cut. It force-disarms immediately and bypasses the flight controller’s in-flight safety checks, which normally refuse a disarm while airborne. An aircraft in the air will stop its motors and fall.Use it when the alternative is worse than the aircraft falling where it currently is. Do not use it as a quicker way to disarm on the ground, because it is a habit that transfers into the air.
A normal disarm is the safe command. It goes through the FC’s checks, which is exactly why it will refuse while the aircraft is flying. Two more gates worth knowing:
  • Fleet-wide arming needs a separate arm-the-broadcast step. Broadcast controls stay inert until you arm the broadcast, and that arming expires by itself after 5 seconds.
  • Extensions can never arm the aircraft. Arm and disarm, motor test, calibration, and reboot are hard-blocked in the Extension command path regardless of declared permissions. Extensions may command takeoff, land, and RTL only, with takeoff altitude clamped between 1 and 120 m.

What the agent refuses while armed

The agent runs an armed interlock in front of its REST API. While the vehicle is armed, every action that can cut the command, telemetry, video, or RC link, or change how the aircraft flies, is refused with HTTP 409:
The interlock covers:
  • rebooting the companion and restarting the supervisor;
  • restarting a unit that carries a link (every ados-wfb* unit, MAVLink, video, RC, mesh, access point, uplink, and cloud units), directly or through a peripheral action;
  • hopping the radio channel, changing transmit power, unpairing a radio, and opening a local radio bind;
  • changing the node’s profile or ground-station role, and a ground-station factory reset;
  • enrolling or disabling MAVLink signing on the flight controller;
  • a flight-controller parameter write (POST /api/params/{name}) and an RC-module parameter write;
  • joining, leaving, or forgetting a Wi-Fi network, re-tagging an adapter’s MAC, and the ground station’s access point, uplink-share, client, Ethernet, and mesh writes;
  • installing, enabling, disabling, or removing a plugin, because a plugin may feed pose or setpoints to the flight controller;
  • POST /api/v1/setup/apply when the body carries a profile, wfb, network, or regulatory section;
  • PUT /api/config whose key is agent.profile, ground_station.role, video.wfb.channel, video.wfb.tx_power_dbm, video.wfb.mcs, or network.regulatory.
The interlock fails closed. If the agent’s vehicle state is missing or older than 3 seconds, the vehicle is treated as armed. It applies to every caller, including requests relayed over the radio. Profiles with no vehicle (no MAVLink router) are not guarded. To act anyway, send the override: "force": true in the JSON body, or ?force=1 on the query string for a request with no body. Use it only when you have decided the action is worth the risk to an armed aircraft. Mission Control adds its own layer on top. Most parameter panels stay usable while armed, with a banner warning that writes go live and a confirmation at save time. A small set is disabled while armed and connected: motor test, ESC and servo output configuration, DShot commands, PX4 actuator test, and frame class. Sensor calibrations (accelerometer, gyro, compass, level, airspeed, barometer, ESC and compass-motor) never start while the vehicle is armed or its arm state is unknown. The interface blocks and the agent interlock are independent; neither depends on the other.

One parameter the agent refuses outright

The agent will not write SYSID_THISMAV, or any other parameter that changes the vehicle’s own MAVLink identity. The write is refused with a reason rather than accepted. The agent addresses the vehicle at system id 1. Changing the vehicle’s system id would make it unreachable by the very connection needed to change it back, and would also stop arm, disarm, mode, and RTL commands reaching the aircraft. There is no way to undo it from the agent afterwards. If you need a non-default system id, set it with a parameter tool connected directly to the flight controller over USB. The receiving side does not treat silence alone as a fault. Video can be quiet for good reasons. It acts when video is silent and it has not heard the drone’s presence beacon recently, which is the combination that means the link is gone rather than idle. Recovery goes in that order: it first tries the channel the peer last announced, and only then sweeps the band. Both ends return to the configured home channel on link loss, so the two sides meet at a known place instead of hunting for each other. The transmit side runs its own check. A transmitter whose byte counter has stopped moving while it is still being fed is treated as dead and restarted, because a process that is running is not the same as a process that is transmitting. Link state is reported as one of disconnected, unpaired, auto_pairing, binding, connecting, connected, degraded, or rf_unverified. The last one deserves attention: the radio is transmitting but nothing has confirmed reception. A transmitter with no listener looks identical to a healthy transmitter from the transmit side alone, so it is reported as unverified rather than shown as connected. On a ground station, the internet uplink fails over in order: Ethernet, then WiFi client, then cellular, then USB tether. A link has to fail three consecutive checks before it is dropped and pass three before it is trusted again, with a 30 second cooldown between switches, so a flapping interface does not thrash the routing table. This ladder carries the ground station’s internet traffic. It does not carry the aircraft’s radio link, and there is no failover from the radio to cellular. If the radio link to the drone is down, the drone is out of contact regardless of the ground station’s internet. If the network you administer the box over goes dead, a guardian climbs a repair ladder without rebooting: re-assert the regulatory domain, renew DHCP, reconnect WiFi, bounce the interface, restart the network backend. It is on by default, and stays on if its configuration is missing or malformed, so that a missing config does not silently disable the thing that recovers the box. If a wired primary link stays down and onboard WiFi has a usable path, the box can fall back to a heartbeat-only reach-back. You keep seeing the node in Mission Control, and video and full telemetry show as unavailable, because that path carries status and not the data plane.

Next steps

Operating limits

What has been measured on ADOS hardware, and what has not.

Regulatory

Your obligations as the operator, and pinning a radio region.

Cockpit and Skills

The Skill Bar, confirmation gestures, and hotkeys.

Geofences and rally points

Draw the boundaries the flight controller enforces.
Last modified on October 11, 2026