Skip to content

Bebop 2 firmware 4.7.1: what is inside

Read from the official update image (firmware/bebop-2-family_v4.7.1.plf, SHA-256 6fe03e1e…ffd2f), unpacked with tools/third_party/extract_plf.py inside a --network none container. Build id ardrone3-milos-4.7.1, built 2018-06-06. Nothing here came from the aircraft itself.

Terminal window
docker run --rm --network none --user "$(id -u):$(id -g)" \
-v "$PWD/firmware/bebop-2-family_v4.7.1.plf:/in/fw.plf:ro" \
-v "$PWD/tools/third_party/extract_plf.py:/in/extract_plf.py:ro" \
-v "$PWD/firmware/extracted/bebop-2-4.7.1:/out" \
python:3.13-slim-bookworm python /in/extract_plf.py /in/fw.plf /out

Run it in a container. The extractor rejects .. but follows symlinks it created earlier in the same run, and the rootfs has 11 absolute ones (var/lib -> /data/lib, home -> /data/home, …). This image routes no entry through them; a different image could write outside the output tree.

Handlers live in /bin/onoffbutton/, named by press pattern:

Presses Script Effect
1 shortpress_1.sh shutdown
3 shortpress_3.sh stop Dragon (stock autopilot) and start ArduCopter, if one is installed
4 shortpress_4.sh shpoison_cli --start_debug: starts telnetd and adbd
long longpress_0.sh toggle Wi-Fi band 2.4/5 GHz; motors chirp as confirmation
very long verylongpress_0.sh factory reset

Telnet (/usr/sbin/telnetd -l /bin/login.sh) is a disabled boxinit service. /bin/login.sh is just exec /bin/sh -l: no password prompt, a root shell (etc/passwd: root, home /data/home/root). With the open Wi-Fi, anyone in range gets unauthenticated root while it runs. 4 presses or plugging in an ASIX USB-Ethernet adapter (asix_setup.sh, which serves DHCP on 192.168.2.1) are what start it. Neither persists across reboot.

/usr/bin/arducopter.sh prefers /data/ftp/internal_000/ardupilot/arducopter, which is FTP-writable storage, so installing it changes no firmware. The boxinit service starts it as:

arducopter.sh -A udp:192.168.42.255:14550:bcast -B /dev/ttyPA1 \
-C udp:192.168.42.255:14551:bcast \
-l /data/ftp/internal_000/ardupilot/logs -t /data/ftp/internal_000/ardupilot/terrain

MAVLink is broadcast on the drone’s own network at 14550 (GCS) and 14551; /dev/ttyPA1 is the GPS. A reboot returns to Dragon.

ArduPilot still builds this board (libraries/AP_HAL_Linux/hwdef/bebop). The stable build is in firmware/ardupilot/ (gitignored binary): 4.7.1-FIRMWARE_VERSION_TYPE_OFFICIAL, commit dbe79216 (2026-08-31), static ARM EABI5, SHA-256 1f1045e7…0680a.

/lib/firmware/brcm/bcm43526.nvm, the shipped generic file:

epagain2g=0 pdgain2g=7 maxp2ga0=76 maxp2ga1=76 femctrl=1 ccode=0 regrev=0

maxp2ga* is in quarter-dB, so 76 is a configured 19 dBm per chain. These match the “stock” preset in GiannisTDM/parrot-bebop2-rf-lab, whose RF power profile (epagain2g=2 pdgain2g=16 maxp=80) we are not applying; see the README there for its own EIRP caution.

“Colibry” is a configuration namespace read by dragon-prog, not a library: there is no libColibry, libNavdata or libHAL in the rootfs. Config lives in etc/colibry/{common,milosboard,...}. Facts below were found by the peer session bebop-engineer and re-checked here against the files.

  • Estimator sample rate adcFsSet = 198.3642578125 Hz (etc/colibry/milosboard/root.cfg:8-10, default working mode).
  • Motor speed range per motor: 3000 to 12200 RPM (etc/colibry/milosboard/drone.cfg:23-24). Test-BLDC’s CLI accepts up to 16000; that is a tool bound, not the flight range.
  • usr/lib/libCondor.so exports condor::Prediction::* (shift_image, getPrediction, smoothPredictionToZero, …): image-shift prediction for the camera. Its setAttitude string belongs to that, not to the pilot setpoint path.
  • lib/firmware/BLDC.cyacd is the motor controller’s (Cypress PSoC) own firmware image: header 1E0680690300, 89 rows of 256 bytes, every row checksum is the two’s complement of its byte sum (Cypress “basic summation”; 89/89 match, a plain sum matches 1/89).

Unverified, from peer notes: an 18-state Kalman filter inside dragon-prog (reverse-engineering/colibry-fusion.md); PSoC command bytes, CRC and loop rate (reverse-engineering/*-DRAFT*.md, marked unverified by their author).

The 4.7.1 image is a family image: etc/colibry/ holds chimeraboard, common, milosblack, milosboard, milosthermal, mykonos3board, and the set is chosen at runtime from the ro.hardware property (peer: etc/init.d/rcS:261, bin/ardrone3_shell.sh:8). Not yet read from the live unit (needs a shell). Indirect evidence:

  • Bebop 1 image build id is ardrone3-mykonos3-4.0.6 and it ships mykonos3board, so mykonos3 is Bebop 1 hardware. This unit identifies as a Bebop 2, so not mykonos3.
  • Bebop 2 image build id is ardrone3-milos-4.7.1. All three milos sets have motorsRpmMax 12200, so the RPM range holds for any Bebop 2 variant; only the exact board name is open. (mykonos3board 11000, chimeraboard 7100.)
  • The Bebop 1 4.0.6 image also carries milosboard.

BLDC_Test_Bench flags, from the binary’s own help text

Section titled “BLDC_Test_Bench flags, from the binary’s own help text”

Source: strings usr/bin/BLDC_Test_Bench (usage block). Authoritative over the peer drafts, one of which had these wrong.

Flag Effect Risk
-I read BLDC info read-only
-O read observed speed read-only
-r ID / -a read one / all motor settings read-only
-S secure start motors (clears errors) spins motors
-D start with arbitrary direction spins motors
-R r1 r2 r3 r4 set reference RPM per motor spins motors
-s ... + -l/-L/-p build and launch step/pulse speed profiles spins motors
-m, -b, -c, -o, -Z, -F multisine, chirp, CSV replay, motor-order test, PWM sweep, feedforward fit spins motors
-H stop motors
-C clear error flag only; does not stop motors
-N keep turning even if the BLDC detects an error disables a safety stop
-w ID VAL write a motor setting persistent write
-E clear the BLDC EEPROM destructive
-A / -i reverse / set propeller spin directions changes config
-G r g p s BLDC GPIOs: red/green LED, P7 reset, switch enable can reset the SoC
-M m play a “soundtrack” through the motors (the boot chirp) spins motors briefly
-X ms I2C timeout
-P, -V, -W, -f factory bench: HAMEG HMP2020 PSU, KERN440 scale, Futaba S.bus RC n/a on a drone

Not every flag is served by this firmware. The peer’s read of the PSoC dispatcher table in BLDC.cyacd (section 11g of reverse-engineering/HARDWARE_REVERSE_ENGINEERING.md) finds handlers for only 0x02, 0x20, 0x40, 0x4D, 0x60, 0x80, 0x82, 0x8D, 0xA0 and 0xFA. There is no handler for reverse (0x06) or settings read/write (0xC0), so -A, -r, -w and -a appear to hit the unknown-command default on 4.7.1. Untested on hardware.

The I2C link uses an XOR8 byte sum, not CRC-16; the earlier drafts’ CRC-16 claim is withdrawn (peer’s decompile of MTB_BLDC_*, 2026-10-01).

Ghidra cross-check of peer claims (read-only, bebop-main)

Section titled “Ghidra cross-check of peer claims (read-only, bebop-main)”

Queried the peer’s dragon-prog-4-7-1 Ghidra container (port 8192) by functions_get / functions_list / program_info; nothing renamed or written.

  • 8437 functions, ARM:LE:32:v8, image base 0x00010000. Confirmed.
  • ACQCON_colibry_Thread_Handler at 0x0004dd18, body 2104 bytes. Confirmed.
  • BEHAVIOUR_Lib_Colibry_Message_Callback at 0x0006e460. Confirmed.
  • setAttitude appears twice, both thunks: 0x00033b30 (12-byte stub) and 0x0060e440 (1-byte external entry). So dragon-prog does not implement setAttitude; it forwards to a shared library, consistent with the libCondor.so export. The six-float signature is Ghidra’s guess.

Comments

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