Skip to content

This unit, as it reported itself

Source: captures/first-contact.jsonl (2026-10-02 04:14 UTC, 16.3 s, read-only), annotated in first contact capture. Drone on a bench indoors, props off, no other controller connected.

  • Name Bebop2-XXXXXX (also its SSID; the suffix matches the end of the serial).
  • Firmware 4.7.1, hardware HW_03. Simulator assumed HW_05; real is 03.
  • Motor controller firmware 6.7.R.4: major 6, minor 7, Release build, 4 motors (format from the XML description of MotorSoftwareVersionChanged). Cross-check: the peer read the same 6.7.R.4 out of the BLDC.cyacd metadata in the 4.7.1 image, by a wholly independent route. So this unit’s motor controller runs the firmware that image ships, unmodified.
  • GPS receiver firmware 3.01F; GPS hardware string empty.
  • libARCommands on the drone: 4.0.0.4.
  • Serial, CPU id and P7ID were reported; kept out of git. The CPU id starts with P7, consistent with Parrot’s P7 SoC.
  • All six self-tests OK: IMU, barometer, ultrasound, GPS, magnetometer, vertical camera (SensorsStatesListChanged, paired from raw order).
  • Motors: noError, and last recorded error also noError, motor id mask 0. The LED was flashing quickly at the time, which the user guide (p.32) lists as “motor error”. Q: see open questions.
  • Magnetometer calibration not required.
  • Battery 100%, a real report at t=0.09 s, not the pyparrot default.
  • Alert state none; flying state landed; motion state steady.
  • 48 flights, 7946 s total (2 h 12 m 26 s), last flight 20 s.
  • Storage: internal 7338 MB, 16 MB used, 0 photos, 0 videos, 0 crash logs.
  • Video autorecord is on, yet 0 videos: either deleted or recorded elsewhere. Q.
Setting Value Allowed range
max altitude 30 m 0.5 to 150
max distance 2000 m 10 to 2000
geofence (no fly over max distance) off
max tilt 8° 5 to 35
max pitch/roll rotation speed 80 °/s 80 to 300
max yaw rotation speed 13 °/s 10 to 200
max vertical speed 1 m/s 0.5 to 6
banked turn off
hull protection not fitted
outdoor on

Tilt, rotation and vertical speed sit at or near their minimums: a gentle, filming-style profile. Max distance is at its maximum with the geofence off.

  • Return-home delay 30 s; minimum RTH altitude 20 m (range 20 to 150).
  • Preferred home type FOLLOWEE (the follow-me target). Chosen: FIRST_FIX. Availability of all four home types: 0 (no GPS fix). Navigate-home unavailable, reason disabled.
  • Flight-plan components: Calibration 1, Mavlink_File 1, TakeOff 1, GPS 0. Flight-plan availability 0 because of GPS alone.
  • Mavlink_File = 1 means a valid file is present: /data/ftp/internal_000/flightplans/flightPlan.mavlink. Q: contents.
  • Last move-to status DONE, all coordinates 500 (none).
  • Horizontal FOV 78°. Virtual gimbal: pan ±35°, tilt −83° to +17°; default and current tilt −13°. Max pan/tilt velocity 20 °/s.
  • Recording 1080p, streaming 480p, 29.97 fps, mode quality, stabilization roll and pitch.
  • Photo format snapshot (16:9 crop), not full fisheye JPEG or raw.
  • White balance auto, exposure 0 (−1.5 to 1.5), saturation 0 (−100 to 100), timelapse off.
  • Video streaming disabled (we did not start it).
  • Anti-flicker auto, but the electrical frequency is fiftyHertz while the country is US (60 Hz). Auto mode uses the frequency “previously sent” by a controller; none sent this session. Q.
  • Country US, automatic. Outdoor Wi-Fi settings on.
  • 2.4 GHz, channel 10, selection auto_2_4ghz.
  • Security open, key empty. Anyone in range can join; the drone accepts one controller at a time.
  • RSSI −23 to −20 dBm at about 1 m.
Group Rate
attitude, speed, altitude, camera orientation about 5 Hz
GpsLocationChanged 2 Hz
PositionChanged 1 Hz
Wi-Fi RSSI 0.8 Hz
sensor-health list resent about every 2.5 s
  • Attitude at rest: roll +3.74°, pitch −0.76°, yaw −75.8°. Roll noise about 0.06° peak to peak. Q: is the bench tilted, or is this an offset?
  • Without a GPS fix, GpsLocationChanged_*_accuracy (i8) wanders across −124 to 119 instead of the documented −1. Positions read 500.0 (unavailable).
  • No date or time events. FreeFlight sets the drone clock on connect; we do not. Video start/stop timestamps read 7 ms after the epoch.
  • Events the XML marks as other products’ still arrived: ConnectedAccessories (an empty list, flags 0x04) and AlertSound.

Read through mcbebop’s list_files, anonymous FTP, no button press needed.

port 21 /data/ftp
internal_000/
Bebop_2/{academy, media, navdata, thumb}
Debug/ gps_data/ log/ lost+found/ scripts/
flightplans/ (empty)
port 61 /data/ftp/internal_000/flightplans (empty, same directory)
ckcm/ckcm.bin 445375 bytes
netdata/ (empty)

log/, scripts/ and netdata/ are empty on this unit.

ckcm.bin is the drone’s system log, and it answers the board question

Section titled “ckcm.bin is the drone’s system log, and it answers the board question”

Not camera calibration, as the name suggested. It is a ulogcat binary log holding the drone’s kernel and userspace messages since boot, and it is live: 445 KB when listed and 560 KB when fetched a minute later. Tags present, by volume: COMMANDS (1561), KERNEL (580), NETMON (150), NETWORK (110), NETWORK_DISCOVERY (67), SETTINGS (52), BEHAVIOUR_COMMANDS (26), TEMPERATURE (20).

This unit runs milosboard. Stated three ways in its own log:

My hardware is milosboard
Product variant : milosboard
Successfully read config from /etc/colibry/milosboard/root.cfg
using include path /etc/colibry/milosboard/default
Machine: Milos board

That settles “which Colibry set does this drone load?” without a shell, and confirms the inference from the per-board masses (milosboard 0.500 kg matching Parrot’s 500 g figure). So motorsRpmMax 12200 and the rest of etc/colibry/milosboard/ are this aircraft’s actual values, not a guess.

Kernel: Linux 3.4.11+ SMP PREEMPT, built 2018-06-06, ARMv7 (412fc097), which matches the firmware build date in build.prop.

The log also carries Mykonos3 board : exporting HSIS to userspace and Mykonos3 board : init done, from the Bebop 1 platform, on a machine that identifies as Milos. Answered by the peer session from the kernel inside the firmware’s partition blobs: those two strings are literal in the 3.4.11+ kernel (built 2018-06-06, matching this log), alongside a separate Milos board : exporting GPIOs line. So the GPIO export has its own Milos message while the HSIS and init-done messages carry the older name unconditionally. Shared code announcing itself under the platform it was written for, not a sign of anything about this aircraft.

The log is operational data (network and discovery records among it), so it stays out of git; captures/ is gitignored in both repos.

Answering the open question from first contact. The drone reports

MavlinkFilePlayingStateChanged_filepath /data/ftp/internal_000/flightplans/flightPlan.mavlink
MavlinkFilePlayingStateChanged_type flightPlan
MavlinkFilePlayingStateChanged_state stopped

and that path reads as though a mission were stored. It is not: the directory is empty over both FTP routes. The filepath field is where a flight plan would live, not evidence that one does. state is stopped, and FlightPlanState.AvailabilityStateChanged is 0, both consistent with no file.

Earlier these notes recorded Mavlink_File = 1 in the flight-plan component list as “a valid file is present”. That reading was wrong; the component being OK is not the same as a mission existing.

Opened with four short presses on the power button; port 23 opened, and 9050 (ADB) with it. Everything below was read through mcbebop’s shell_read, which allows eleven read-only command names and refuses shell metacharacters. The shell disappears on the next reboot.

The radio NVM is the generic file, not a per-unit calibration

Section titled “The radio NVM is the generic file, not a per-unit calibration”

This answers the open question. /lib/firmware/brcm/bcm43526.nvm on the live aircraft is identical to the copy in the firmware image: epagain2g=0, pdgain2g=7, maxp2ga0=76, maxp2ga1=76, femctrl=1, ccode=0, regrev=0.

It is a Broadcom reference template that was never specialised per unit, and two things in it prove that rather than merely suggest it:

  • boardnum=${serno} is present literally, an unexpanded shell variable.
  • Its macaddr carries a Broadcom-registered prefix, while this drone’s actual BSSID is in Parrot’s A0:14:3D range. The MAC in the file is a placeholder.

The file also has an #[NVRAM ONLY] section and otpimagesize=600, which fits the usual Broadcom arrangement: the per-unit calibration lives in the radio chip’s one-time-programmable memory and the NVM file only supplies board-level defaults. So there is no per-unit calibration file to read, and comparing a unit’s NVM against the image will always match.

bcmwl nvram_dump answers bcmwl: Unsupported on this firmware, so the runtime values could not be read back for comparison. Q.

Per-unit hardware revision, and the board’s GPIO map

Section titled “Per-unit hardware revision, and the board’s GPIO map”

/sys/kernel/hsis/ is exported by the kernel and holds this aircraft’s own hardware descriptors. Revisions first:

hwrev 14 pcbrev 3 p7rev 3

The rest is the board’s GPIO map, each entry giving the pin number the SoC drives. Read in full over ADB:

entry pin what it is
camera_h_mclk / camera_h_pwdn 11 / 132 forward fisheye camera clock and power-down
camera_v_mclk / camera_v_pwdn 9 / 129 downward camera, the one the self-tests call vertical_camera
clkin_gyro, fsync_gyro_p7, gyro_int_p7 8, 90, 91 gyro clock, frame sync and interrupt
magneto_int_p7 124 magnetometer interrupt
gps_power_en 81 GPS power enable
reset_psoc 199 reset line to the motor controller
reset_wifi 9 radio reset
fan, thermal_pwm 85, 6 cooling
super_led_pwm 0 the status LED
user_on_off 201 the power button, the one the press patterns read
host_mode_3v3, host_mode_on, select_alim_us, usb0_oc 202, 203, 200, 204 USB host mode, supply select, over-current

The pin map was predicted from the kernel image before it was read, and all 19 predictions matched. The peer session decoded mk3_hsis out of the 2018 kernel and published its expected values; every one agrees with this aircraft. That makes the map a confirmed decode rather than either side’s inference, and it confirms the Milos entry routine patches super_led_pwm from -1 in the image to 0 at boot.

Its revision rule, hwrev = pcbrev + 11 for Milos, also fits: this unit reads pcbrev 3, hwrev 14. One unit fitting a rule is consistency, not proof; it becomes evidence when a second Bebop 2 with a different pcbrev reads the predicted hwrev.

The discrepancy that chasing turned up, now answered: the attribute table has 22 entries pointing at 22 distinct struct cells, but the struct holds further cells with no file. The peer read two of them out of the user_gpio export table as GPS_FLASH (+0x50) and RESET_GNSS (+0x54, set to 83 on Milos), plus one cell at +0x0c that nothing references at all. So the struct and the published attributes are genuinely different sets that both happen to number 22, and reset_gnss is set at boot but never exposed.

Worth recording for anyone reading that table: its entries are 12 bytes (field pointer, flags, name pointer) in two copies, and the name follows its field pointer rather than preceding it, which makes neighbouring entries look off by one if read the other way round.

A caveat the peer raised against its own earlier claim: only seven distinct GPIO names live in that export table, stored twice (one copy per board, differing in a flags word), not fourteen. The other seven (FAN, RESET_PSOC, MAGNETO_INT_P7, HOST_MODE_3V3, USB0_OC, CAMERA_V_PWDN, CAMERA_H_PWDN) are registered through a helper with the number passed in a register, so which hsis field feeds each of those is an agent’s reading rather than a verified decode. Their values are confirmed on hardware either way; it is the field-to-name binding for those seven that is softer.

Two cameras, which matches the self-test list. reset_psoc on a kernel-exported pin means the main SoC can hold the motor controller in reset directly, which is worth knowing next to the BLDC work: it is a path to the motor controller that does not go through I2C at all.

getprop does not exist on this firmware and is not on the shell’s PATH under any of the usual names, so Android-style property reads are not available; /etc/build.prop must be read with cat instead. A long command line is wrapped by the pty across two lines, which the tool’s output handling survives but which is worth knowing if anything ever parses the echo.

Settings as the aircraft reports them (2026-10-02, live)

Section titled “Settings as the aircraft reports them (2026-10-02, live)”

Read for the peer session’s cross-check. Note current is how this unit happens to be configured; min/max are the bounds the firmware allows, and the bounds are what a config file should be compared against.

setting current min max unit
MaxTiltChanged 8 5 35 deg
MaxVerticalSpeedChanged 1 0.5 6 m/s
MaxRotationSpeedChanged (yaw) 13 10 200 deg/s
MaxPitchRollRotationSpeedChanged 80 80 300 deg/s
MaxAltitudeChanged 30 0.5 150 m
MaxDistanceChanged 2000 10 2000 m

NoFlyOverMaxDistanceChanged = 0 (geofence off), BankedTurnChanged = 0, HullProtectionChanged = 0 (no hull fitted), OutdoorChanged = 1.

Two different tilt quantities, which is why they disagree

Section titled “Two different tilt quantities, which is why they disagree”

The camera figures and the piloting figures are unrelated, and conflating them is what made the earlier “tilt range mismatch” look like a contradiction:

  • MaxTiltChanged is the pilot authority limit: how far the aircraft may lean. 8° here, settable 5 to 35.
  • CameraSettingsChanged tilt −83 to +17 with pan ±35 and FOV 78 is the virtual gimbal, a crop window over the fixed fisheye. Confirmed live by sending Camera.OrientationV2(tilt=15) and reading 14.999999 back.
  • videoTiltRange [-75, 30] in etc/colibry/common/video.cfg is a third quantity again and governs neither of the above directly. Q: what does it bound?

Connecting with FreeFlight Pro changes settings on the aircraft

Section titled “Connecting with FreeFlight Pro changes settings on the aircraft”

Found 2026-10-02 by comparing telemetry before and after the app-side session connected FreeFlight Pro 5.2.7 to the drone. That session navigated settings screens, declined the firmware update, flew nothing and believed it had changed nothing. Three values changed anyway:

setting before after
NoFlyOverMaxDistanceChanged_shouldNotFlyOver 0 (geofence off) 1 (on)
ReturnHomeDelayChanged_delay 30 s 3 s
BankedTurnChanged_state 0 1

These were read from the aircraft before the app ever connected and read again afterwards, so the comparison is against the drone’s own reports rather than against expectation. A reboot alone does not explain it: these persist in /data/dragon.conf, which survives power cycles, and the other settings in the same file (max altitude 30, max distance 2000, max tilt 8, vertical speed 1, rotation speeds, hull protection, outdoor, autorecord, country) are unchanged.

Settled 2026-10-03, after one wrong turn. A peer session decompiled FreeFlight Pro 5.2.7 and found the mechanism in code: on connect it saves its own settings copy (or its defaults if the store is empty), reads the drone’s state, then replays the saved values back to the drone. The replay list includes all three that changed, and the app’s built-in defaults match two of the observed values (return-home delay 3 s, banked turn on). A fresh install writing its defaults over the aircraft is exactly the behaviour we saw.

I had claimed the app’s second session sent no setting commands at all, because I searched the log for Cmd Rx lines and found only date, time and two state requests. That was the wrong pattern. The drone does not log a setting write as Cmd Rx. It logs the effect, under BEHAVIOUR_SETTINGS_PLATFORM, as a ..._MSG_SET_... line. Searching for those shows the app writing settings two to four seconds into the same session I had called clean:

[+2.41] Return Home delay after disconnection set to 3
[+2.45] BEHAVIOUR_FEATURE_SETTINGS_MSG_SET_NO_FLY_OVER_MAX_DISTANCE : 1
[+3.77] BEHAVIOUR_FEATURE_SETTINGS_MSG_SET_MAX_DISTANCE : 2000.000

Note the GET_ lines at +0.44 and the SET_ lines at +2.45: the first is the drone reporting its own state in answer to “all settings”, the second is the app writing. Reading only the Cmd Rx lines hides the writes entirely.

So the original claim was right after all: the app pushes its stored preferences on connect, every connect. It is now backed by the app’s own code and by the drone’s log, rather than by one before-and-after reading.

It also sets the drone’s clock, which nothing else we have used does. The aircraft has no battery-backed RTC and boots at the epoch, so before this its timestamps were really uptimes. After an app connects, they are wall-clock. That matters for anything reading file times or log stamps: whether they mean anything depends on whether a phone has been near the drone since it booted.

Practical consequences:

  • “I only looked at screens” is not the same as “I changed nothing” for any app that syncs preferences. The only way to know is a before-and-after read.
  • Two of the three changes are, by luck, in the safer direction: the geofence is now on, which this unit had switched off against a 2000 m limit. The return-home delay dropping from 30 s to 3 s is a behaviour change worth knowing before a flight, as it decides how long a lost link is tolerated before the aircraft starts flying home on its own.
  • One transcription caution: the app-side report listed “max altitude 2000 m, max distance 30 m”, which is the pair swapped. The aircraft reports altitude 30 and distance 2000, both before and after, so that was a reading error on the app side rather than a change. Values that come back as a plausible-looking swap deserve a second look.

Battery note: immediately after the swap the aircraft reported BatteryStateChanged_percent = 0 on a freshly fitted battery, refreshing every few seconds. Most likely the gauge had not yet characterised the new pack. Worth re-checking rather than assuming, since a genuine 0 would matter. Q.

Settings restored, and one deliberately not restored (2026-10-02)

Section titled “Settings restored, and one deliberately not restored (2026-10-02)”

Ryan asked for the app’s three changes to be reverted and the max distance tightened. Carried out through mcbebop, armed for 5 minutes with the reason recorded, then disarmed. All values below were read back from the aircraft afterwards, not assumed.

setting original app set now
return-home delay 30 s 3 s 30 s restored
banked turn 0 1 0 restored
max distance 2000 m 2000 m 150 m tightened
geofence 0, off 1, on 1, on deliberately left on

Why the geofence was not reverted. The request was to revert all three and tighten the distance, but those two instructions contradict each other on this aircraft. Parrot’s own wording for NoFlyOverMaxDistance is “1 if the drone can’t fly further than max distance, 0 if no limitation on the drone will be done”. So with the geofence off the distance limit is not enforced, and tightening it from 2000 m to 150 m would have changed nothing that the aircraft acts on. Reverting the geofence would have undone the point of the tightening.

So the end state is the original configuration except for two deliberate differences: the distance limit is 150 m rather than 2000 m, and it is actually enforced. Turning the geofence back off is one command if that is wanted; it would make the 150 m advisory only.

Unchanged and verified: max altitude 30 m, max tilt 8°, vertical speed 1 m/s.

Battery still reads 0% on the fresh pack, now hours after the swap and refreshing every few seconds, so the earlier guess that the gauge merely needed time to characterise the new battery looks wrong. Either the pack is genuinely flat, or this aircraft does not report a meaningful percentage for it. Worth checking against the drone’s own LED and against FreeFlight Pro, which showed 100% earlier in the same session. Q.

Comments

Spotted a mistake, or tried this on your own Bebop? Leave a note. Comments are kept on this site's own server.