Colibry sensor fusion, Bebop 2 firmware 4.7.1 (static analysis)
Rootfs: firmware/extracted/bebop-2-4.7.1/rootfs Ghidra: usr/bin/dragon-prog (ARM32, image base 0x10000), read-only use, port 8192. Labels: EVIDENCE = file + string/address. INFERENCE = my reading, not proven.
1. Where the code lives
Section titled “1. Where the code lives”- EVIDENCE: etc/build.prop has
ro.parrot.build.variant=milos, uidardrone3-milos-4.7.1. So the Bebop 2 config set is etc/colibry/milosboard/ layered on etc/colibry/common/. - EVIDENCE: no libColibry, libNavdata or libHAL exist in usr/lib or lib.
grep -rl 'bruitAcc|PinitVit|estimation.P.0.0'over the whole rootfs hits only usr/bin/dragon-prog and etc/colibry/common/estimator.cfg. The estimator is statically linked into dragon-prog (C++, Eigen-based). - EVIDENCE: libCondor.so is NOT the attitude estimator. Its symbols are condor::Engine, Prediction, System33, AdaptiveFrameRateFSM, MotionEstimate, Camera, DumpWriter, with
Engine::setAttitude(float x6),Compute,getTranslation, source pathpackages/Condor/Sources/motion/CONDOR_Prediction.cpp. It is the vertical-camera image-registration (optical flow) engine. It consumes an attitude and produces translation in pixels. The “setAttitude thunk” in dragon-prog feeds it attitude, it does not produce it. - EVIDENCE: libHAL is not a library here; the HAL is compiled in (strings
HAL_IMU_Temperature_Control_init,HAL_ultrasound_launch_acquisition,_HAL_gyro_SI_t,_HAL_fifo_gyro_SI_t,_HAL_magn_mG_t,_HAL_gps_data_t). - EVIDENCE: class names in dragon-prog (Ghidra classes_list):
Mykonos3::EstimatorManagerImpl(dtor 0x001628a0),Mykonos3::EstimatorStateInit / EstimatorStateError / EstimatorStateRunningFlying,Colibry::IMUSensorProcessor, BaroSensorProcessor, MagnetoSensorProcessor, GPSSensorProcessor, XsensProcessor, BatterySensorProcessor, SensorBLDCProcessing, ESCMotorModel::Estimator,Features::CommonEstimation,Features::GlobalNavigationEstimation. Namespace “Mykonos3” is the shared platform code name, also used for Bebop 2 (INFERENCE, since wai_is_mykonos3 exists at 0x00033ef0 and the milos config overrides it).
2. Architecture as seen in the binary
Section titled “2. Architecture as seen in the binary”2.1 Acquisition loop
Section titled “2.1 Acquisition loop”- EVIDENCE: ACQCON_colibry_Thread_Handler @ 0x0004dd18. Loop
while (*param_1 != 2): waits on a ring buffer (FUN_004407f8, timeout 100000),memcpys one 0x9f0-byte frame, fixes up two internal pointers inside the frame (+0x138, +0x13c choose between frame+0x48/+0xc0 and +0x84/+0xfc, i.e. double-buffered sub-arrays), then calls a table of callbacks stored in the thread struct (param_1[0xbd], [0xbe], [0xc1], [0xbf], [0xc0], [0xc2], [0xc3]). It also checks the 64-bit timestamp at frame+8 for gaps (compares with a delta; 500000 us threshold, logs a warning and calls FUN_0004d568 on a jump). It does not itself contain filter math. - EVIDENCE: root.cfg/
adcFsSet:adcFsis 198.3642578125 Hz (“default” and “fvt”) and 200.0 in the simulator. root.cfg:adcTs = adcTs/adcFsat runtime. Process-noise comments in estimator.cfg useadcTs, so the filter runs at about 198 to 200 Hz, one step per frame. (Your “~90 Hz” figure for the loop is not supported by config; see open questions.) - EVIDENCE: helper.cfg: IMU filter comment says “fs = 1000 Hz” (4th-order Butterworth, fc = 30 Hz;
filterImuBCoef/ACoef,fifoSize = 5). INFERENCE: the IMU is sampled at 1 kHz by HAL and a 5-sample FIFO per frame is decimated to ~200 Hz.
2.2 Estimator step (the actual fusion)
Section titled “2.2 Estimator step (the actual fusion)”EVIDENCE chain, all in dragon-prog:
- Vtable-style slot table at 0x00578650 holds FUN_0022cd34 (a per-frame “process all” function; only reference is a DATA xref). It calls a sequence of per-subsystem steps on singleton objects (FUN_00103610 / FUN_00103740 / FUN_00103820 return singletons), and in the middle calls
FUN_0037d884(singleton_00103610 + 0x538). - FUN_0037d884 (call this the GlobalEstimator step, INFERENCE for the name;
GlobalEstimatorappears in root.cfg logSet byNode). Its body is a fixed sequence on the estimator object: nine or so “prepare/predict” calls (FUN_00354940, 00355024, 00354014, 00354474, 0035809c, 00378150, 00357d50, 003554a8), then a branch on a config byte (**(char**)(this+0x110), INFERENCE: GPS/global-navigation-enabled) selecting between two groups of observation builders (FUN_00355600/003563f0/0035642c versus FUN_00355dc0/00356328/00356180), thenFUN_003553b0(this,0),(this,1), and(this,2)only if**(char**)(this+0x208), then more observation builders (00355454, 00356800, 003552f8, 00356474, 003564a4, 003564d8, 00356598, 00356654, 00356760), then FUN_0037c0f8 (the measurement update), then post steps (00354bc8, 00355110, 00356c90). - FUN_0037c0f8 (measurement update): EVIDENCE it computes the observation count as a weighted sum of byte flags at this+0x1298..0x12aa and 0x1404 (weights 1, 2 or 3 per flag, plus
*(int*)(this+0x360) * flag(0x129f)), then allocates aniVar11 x 18(0x12) float matrix with posix_memalign (iVar11 * 0x48bytes, i.e. rows x 18 x 4). That is a dynamically sized stacked observation matrix H with 18 columns. INFERENCE: this is a single-shot stacked EKF/Kalman update where each sensor observation contributes 1 to 3 rows only when its validity flag is set. - FUN_0037a16c (Kalman gain; called only from FUN_0037c0f8 @ 0x0037c5fc). EVIDENCE: uses Eigen dynamic matrices, 18-wide, with a Cholesky/LLT-type SPD check; on failure logs the string at 0x004e195c: “Kalman Gain computation failed : innovation covariance matrix is not symetric definite positive”. Structure (decompiled page 1 of 3): compute
H*P(FUN_00377d5c with n=0x12), then(H*P)*H^T(FUN_00377680), add to R (elementwise float add loop, FloatVectorAdd) giving the innovation covariance S, factorise S, then solve for K. INFERENCE: K = P H^T S^-1 via LLT; I did not read pages 2 and 3, so the exact state-update and Joseph-form-or-not question is open. - EVIDENCE: the exported debug variables
estimation.P.0.0 ... estimation.P.17.17(18 diagonal entries of P, strings in dragon-prog) confirm an 18-state error covariance. The 18 columns in the update and the 18 P-diagonals agree.
2.3 What the 18 states are
Section titled “2.3 What the 18 states are”This is the part I cannot prove. Do not treat the list below as fact.
- EVIDENCE (exported estimation.* names, dragon-prog): accelerometerBias.0-2, attitudeEulerAngles.0-2, barometerAltitudeBias, linearVelocityNED.0-2 / linearVelocityBody / linearVelocityHorizontal, position.0-2, altitude, windNED.0-1, windBody.0-2, airSpeedBody.0-2, magneticDeclination, gamma, Cz (vertical aerodynamic model:
vertical_model_Gamma,vertical_model_Cz), plus separate gyro bias exportsgyro_bias_x/y/z_rad_s,acc_bias_x/y/z_m_s2. - EVIDENCE (estimator.cfg init-covariance names, which map one-to-one to state groups):
PinitVit(velocity, mm/s),PinitAngle(attitude, deg),PinitAlt(altitude, mm),PinitHorizontalBias/PinitVerticalBias(accelerometer bias, mg),PinitDerive(gyro drift, deg/s),PinitBP(baro pressure bias, Pa),PinitCap(heading, deg),PinitPosition,PinitVent(wind). - INFERENCE: a plausible 18-state layout from those groups is attitude error (3) + gyro drift (3) + velocity (3) + position (3, only when
alwaysIntegratePosition/GPS) + accel bias (3) + baro bias (1) + wind (2). That sums to 18 but I have no direct evidence for the ordering or that exactly these are in the vector; the observation-flag weights (1/2/3 rows) are the only tie to dimension. A way to settle it: decompile FUN_00354940/00355024 (likely build F and Q) and read which P index eachPinitXconfig writes. - Separate smaller filters exist and are NOT part of the 18-state one:
userTrajectoryX.0-5/userTrajectoryP(6-state follow-me target tracker, config blockuserTrajectoryConfig), battery model Kalman (batteryModel*params; statesnoLoadBatteryVoltageand beta), aerodynamic torque estimator (aerodynamicTorque*,AerodynamicTorqueEstimator), vertical model (verticalModel*, Gamma/Cz recursive estimator), magnetometer calibration/heading filter (noiseMagnetoMeasurement,noiseProcessMag,magneticFieldCovInit,magnetoBiasCovInit).
3. Sensor inputs, rates, validation and gating (EVIDENCE from etc/colibry/common/*.cfg unless noted)
Section titled “3. Sensor inputs, rates, validation and gating (EVIDENCE from etc/colibry/common/*.cfg unless noted)”IMU (ICM-20608; gyro + accel)
Section titled “IMU (ICM-20608; gyro + accel)”- 4th-order Butterworth low-pass fc 30 Hz at fs 1000 Hz (
filterImuB/ACoef, helper.cfg). Gyro is integrated to a cumulative quaternion with a delay line (GyroCumulQuat,GyroIntegrated,GyroUnbiasedFifo,filterQuaternionDelayedGyro*,filterIntegratedDelayedGyro*with an 11-tap pure delay). INFERENCE: the delay aligns gyro integration with the camera frame time for CONDOR (vcam about 66.12 fps,vcamFrameRateSet). - Gyro bias persisted in
/data/gyro_calibration_v1.bin(pathToGyroBiasFile), saved only if bias changed by more thanminBiasDiffToSaveGyroCalib = 0.18 deg/sand IMU temperature aboveminTemperatureToSaveGyroCalib = 49.5 C. - Accelerometer bias file
/data/acc_calibration_v1.bin(pathToAccBiasFile). - IMU thermal compensation: files
/factory/Thermal_IMU_bias.fact.csv,/factory/Thermal_IMU_ortho.fact.csv, user overrides/data/Thermal_IMU_bias.user.csv,/data/Thermal_IMU_ortho.user.csv(strings),ProcessedImuTemperature, temperature derivative filterfilterTemperatureB/ACoef; the IMU is heated and temperature-controlled (_apply_heating_temperature_reference,HAL_IMU_Temperature_Control_init). Calibration binary: usr/bin/AR_Drone3_IMU_calibration. - Dynamic inclinometer gating (accel as a tilt reference):
normAccMin 0.015,normAccMax 0.5,bruitInclinoMin 5 deg,bruitInclinoMax 30 deg,reductionBruitInclino 5. INFERENCE: tilt-measurement noise is scaled between 5 and 30 deg depending on how far |accel| is from 1 g. - Aerodynamic observation:
RAero 200(x MG)^2,normAccHorMax 1 g,timeBeforeObsAcc 200. INFERENCE: horizontal accel is used as a drag/airspeed pseudo-measurement via the Cx/Cy drag coefficients (drone.cfgCxShell 0.35,CyShell 0.40,CzShell 0.9). - Sanity tests:
BiasZInclThreshold 45 deg,accHorizontalityTest 30,maxEstimatedHorizontalBias 100 mg,maxEstimatedVerticalBias 350 mg,AccThresholdForAngleInit 3.4. - Init: FSM string
Colibry estimation FSM still not started: immobility detector : %d , baro is running : %d(0x004c593c, logged from FUN_0022c998, once per 5 x some period). So the estimator waits for an immobility detector and a running barometer before leaving init (EstimatorStateInit).PinitDerivehas Default 0.5, Immobility 5.0, Calibration 0.04 deg/s variants. - Cutout (motor stop) detection uses accel and gyro: strings “CutOut detected with accZ and gyro diff”, “…3 axis accelero and diff gyro”,
cutoutDetectionAcceleroThreshold,cutoutDiffGyroThreshold, and angle-error counters. This is a safety monitor on the estimate, not a fusion input (INFERENCE).
Barometer (MS5607)
Section titled “Barometer (MS5607)”barometerFrequency 75 Hz. Low-passfilterPressureB/A(Butterworth 2nd order, fc 0.5 Hz at 75 Hz). Validity filtervalidDataFilterB/A(first-order, fc/(fs/2)=0.07), covariance filtercovBaroFilterB/A.- Range gates
minPressure 30000,maxPressure 120000Pa;thresholdValidPressure 100;thresholdCovBaro 2000 Pa^2;maxCovBaro 200000 Pa^2. - Defective-baro detector, exact log strings: CONSTANT VALUE, COV, DRIFT, GLITCH, HIGH VALUE, LOW VALUE, NO MEASURE RECEIVED, STEP;
baroDefectiveUnvalidDataCount 600; drift testpressureDriftMaxValue 0.085 m/s, buffer 5 s, check duration 30 s. Defect raises COLIBRY_EVENT_MSG_DEFECTIVE_SENSOR_DETECTED and “Animation stopped because of defective barometer”. - Compensation for rotor/airflow effect on static pressure:
frontal/lateralAirspeedVector,...AngleVector,...AltitudeCompensationMap(milosboard/helper.cfg has a non-zero frontal map: airspeed 10 m/s gives -2.8 and +0.22, 30 m/s gives -8.4 and +1.48). Filebarometer_compensation.csvis copied to /var/lib/colibry by etc/init.d/rcS (first boot only) frometc/colibry/<hardware>/, but that source file is NOT present for milosboard in this rootfs. - Observation noise
RPression 0.3(Pa)^2-scale, process noisedBiaisPression 1.0, bias statebarometerAltitudeBias/PinitBP. Ground effect: milosboardheightGroundEffect 0.5 m,durationLandingWithoutPress 5 s.diffAltitudePressMax 8 m,altitudePressMinForTakeOff 0.6 m. - Separate use for follow-me user barometer (
userTrajectoryConfig.barometer*,barometerMaxGapForValidity 60 Pa,barometerValidPressureFilter*).
Magnetometer (AK8963)
Section titled “Magnetometer (AK8963)”- Heading Kalman-style block (
noiseMagnetoMeasurement 60 mG,noiseProcessMag 300,noiseProcessMagBias 50,magneticFieldCovInit 2500,magnetoBiasCovInit 40000). Innovation gating:magnetoInnovationTestSigmaLevel 2.0,magnetoBiasValiditySigmaLevel 4.0,biasUpdateAllowedSigmaLevel 1.0,minMagnetoCalibInnovationRatioToValidateAxis 0.95, bias-covariance thresholds by accuracy (HighAccuracy 180, MediumAccuracy 400/400/900, LowAccuracy 3600/3600/8100 mG^2).invalidCalibrationDetectionDuration 2 s,thresholdTimeToAbortCalibration 30 s. - Bias persisted in
/data/magneto_calibration_v1.bin. Motor-current magnetic compensation exists but is disabled (enableMagnetoCompensation = false). - Heading fusion gains:
headingFusionGainMagneto 0.0015,...Carpet 0.3,...Xsens 0.1. - Sensor enable flag
magnetometer = trueinenabledSensorsSet(root.cfg).estimation.magnetoCalibrationState,magneticDeclinationandmagneticDeclinationLockedare exported.
Ultrasound
Section titled “Ultrasound”ultrasoundFrequency 15 Hz,nbSamplesUS 5,thresholdCoherenceUS 0.15, rejection thresholdsUsRejectionStepThreshold 0.12,...PredictionThreshold 0.12,...NewGroundHeightThreshold 0.07,...FirstGroundHeightThreshold 0.3, presence windowUsAltMinForPresence 2.0 / UsAltMaxForPresence 3.5(m),timeToLossUsSignal 1 s,nbNoUsMeasuresToResetBuffers 67(comment “5 samples at 15 Hz over 200 Hz”, confirming the 200 Hz filter rate). Observation noiseRUs 50(mm).- Modes
UsModeHighPower 0 / Low 1,UsMinValueHigh 0.6,UsMinValueLow 0.4,us_transit_*. Threshold tables: etc/ultrasound/threshold_0.txt and threshold_1.txt, config etc/ultrasound/default-config.cfg. Defective-US detector:ultrasoundDefectiveMaxNoise 0.03,...MinNoise 0.01,...BufferSize 30. altitudeUSSensorBias 0.04 m(US to ground when landed).estimation.useUSis an exported flag.
Vision (vertical camera optical flow via libCondor)
Section titled “Vision (vertical camera optical flow via libCondor)”- Camera about 66.12 fps (
vcamFrameRateSet). Exports:sensors.visionRawData.timestamp_vcam_ns / vision_defined / vision_indicator,sensors.visionVelocity.vision_speed.0-2,estimation.visionAltitude,visionAltitude2,visionFlag,estimation.RVision.0-2,lastReceivedCameraTimestamp. - Gating:
visionIndicatorActive = falsein common (needs checking if overridden elsewhere; milosboard/estimator.cfg does not override it),fusionVisionVz = false,visionIndicatorRejectThreshold 0.8(1.5 when hovering without GPS),cptVisionFail 100,visionSigmaMin 0.15,gpsVisionSensitivity 0.3, speed noise floorvisionSpeedRminSet(Default 0.25, GPS 1.0 in (m/s)^2) plusvisionSpeedRminAltitudeGain 0.0011per m^2,dRVisionSpeedZ 0.3,altitudeVisionSensorToGround 0.04 m(milosboard). Vision speed low-passvisionFilterB/A(tau 0.05 s at 66 Hz) and sigma filter (tau 0.5 s).maxTimeNotVisionS 20. - “Magic carpet” (visual pattern) position/heading observation with dynamic noise
RCarpetHorizontalMin/Max 10/20 cm,RCarpetAltitudeMin/Max 5/15 cm,minNbPatternForEstimation 2;sensors.visionCarpetData.*. - Sensor enable flag
verticalCamera = true. NoteDefective vertical camera detectedstring.
GPS (on ttyPA1; config in etc/gps_config.txt, gps_firmware.bin)
Section titled “GPS (on ttyPA1; config in etc/gps_config.txt, gps_firmware.bin)”dtGps 0.2 s(5 Hz),gpsLag 1(frames of delay compensation in common; mykonos3board uses 3; milosboard does not override, so 1 on Bebop 2),timeToResetPastSpeed 0.4 s.- Fix qualification:
nbrSatelliteMinToFirstFix 7,nbrSatelliteMinToKeepFix 5,thresholdPositionDeviationErrorFirstFixGps 1.7 m,...ToFixGps 2.3 m,thresholdSpeedDeviationErrorToFixGps 0.8 m/s,thresholTimeToValidGpsFix 4 s,thresholdTimeToLostGpsFix 1 s,thresholdTimeWithoutNewGpsData 2 s,gpsAltDeviationErrorMax 2 m,gpsInvalidCoordinates 500. - Noise:
RSpeedGPS 0.2(m/s),RPositionGPS 2.0(m).fusionVzGps = falsecommon,trueon milosboard (so Bebop 2 fuses GPS vertical speed). Events COLIBRY_EVENT_MSG_GPS_OK / GPS_NOK. estimation.lastGPSHorizTimestamp,gpsSpeedBody.*,userGpsPosition.*exported.
External “Xsens” fusion (dev rig)
Section titled “External “Xsens” fusion (dev rig)”xsensFusion { pos = true; speed = false; angles = false; },xsensInitDuration 0.5 s,XsensProcessor,XsensFusion. INFERENCE: an external mocap/INS input for engineering, not used on a retail Bebop 2.
Sensor selection and health
Section titled “Sensor selection and health”enabledSensorsSet(root.cfg) booleans: ultrasound, verticalCamera, gps, magnetometer, all true in default/fvt/simulator.enabledGlobalNavigationSensors = true.estimation.estimatorSensorFlag(string at dragon-progestimatorSensorFlag) is exported. INFERENCE: it is the bitfield of which observation flags (the bytes around this+0x1298..0x12aa) were active. FUN_003553b0(idx 0..2) writes flag bytethis+0x1299+idxas the AND of: not-flag(+0x1389), optional flag(+0x1388), per-idx bytes at +0x1384+idx and +0x138a+idx, optional +0x1387, and not per-idx byte +0x138d+idx. INFERENCE: three per-axis observation enables (probably the three accelerometer/aero or three vision-speed axes); evidence is only the 3-way structure.
4. Tunables and where they live
Section titled “4. Tunables and where they live”EVIDENCE:
- etc/colibry/common/estimator.cfg (280 lines): all covariances. Layout of the file:
Pinit*(initial P), process noise (bruitAcc 2.5 mg,bruitVol 2.0,bruitVolVert 0.4,bruitAccVol 40,rdw 0.5,rdwSol 5.5,feGyro 0.01,feSol 0.02,dBiais 0.3,dBiaisZ 0.65,dDerive 0.03,dBiaisPression 1.0,dBiaisTh 15,dBiaisMvt 2,dDeriveTh 0.2,dVent 0.02,TauVent 0.5,tauWindToZero 1.5,dPosition 0.5), observation noiseRAero 200, RSpeedGPS 0.2, Rderive 1.0, RbiaisZ 10, RPression 0.3, RUs 50, RPositionGPS 2.0. Comments say actual value is squared and/or scaled byadcTs, e.g.bruitAccbecomes(bruitAcc*dt)^2,dBiaisbecomes(dBiais*MG*sqrt(min))^2*adcTs(a random-walk spec in units per sqrt minute). - Sub-estimators in the same file: vertical model (
verticalModel*), aerodynamic torque (aerodynamicTorque*, 10-coeff initial model, 2nd-order Butterworth 40 Hz at 200 Hz), heading (noise*Mag*,headingFusionGain*), user trajectory (userTrajectoryConfig {…}), battery Kalman (batteryModel*), velocity/vision rejection (gainRejetVisionPos,velocitiesAccGain 0.5,gainEstPos 0.5). - etc/colibry/common/helper.cfg: sensor pre-filters, GPS/baro/US/mag validation. etc/colibry/common/drone.cfg and milosboard/drone.cfg: mass 0.5 kg, Jx 0.001491, Jy 0.001458, Jz 0.003328, motor RPM 3000 to 12200, drag
Cx/Cy/Cz. - Variant overrides for Bebop 2: etc/colibry/milosboard/estimator.cfg (
altitudeVisionSensorToGround 0.04,heightGroundEffect 0.5,durationLandingWithoutPress 5.0,verticalModelInitCz 0.9, 29-point battery curve) and milosboard/helper.cfg (fusionVzGps = true, baro compensation maps, hover power numbers). mykonos3board holds the Bebop 1 overrides (not applicable). - Config loading: files are
@included from milosboard/root.cfg via@include "init.cfg"(selectsworkingModeInitValue = "default"),../drone.cfgetc. (relative to the board dir, which falls back to common via Parrot’s loader; INFERENCE on the fallback order). The binary itself says “Config file read ended with error : %s, exiting Colibry” and “Default parameters setting ended with error”, andcorrectVariablesFromConfigin the config manager post-processes values (squares covariances, multiplies by adcTs) per the comment at the top of root.cfg. The registrations of keys likebruitAccare at 0x004c1338 (3 xrefs: FUN_00365a50, FUN_0022504c, FUN_001988bc), unread. - Runtime state files (RW): /data/acc_calibration_v1.bin, /data/gyro_calibration_v1.bin, /data/magneto_calibration_v1.bin, /data/Thermal_IMU_*.user.csv, /var/lib/colibry/barometer_compensation.csv, /data/colibry_json (string), /data/ftp/internal_000/condor.txt (string). None are in the extracted rootfs (data/ is empty).
- /factory/ is empty in the extraction, so factory IMU thermal calibration is not available here; /calib holds camera tuning only.
5. Open questions
Section titled “5. Open questions”- State vector order. Which of the 18 entries are attitude error, gyro drift, velocity, position, accel bias, baro bias, wind? Need FUN_00354940 / 00355024 / 00354014 / 00354474 / 0035809c / 00378150 (suspected F, Q builders) and the code that writes the
estimation.*exports to pin down indices. The initialPinit*assignments would give the order fastest. - Is attitude kept as a quaternion integrated from gyro outside the 18-state (error-state/multiplicative EKF) or are Euler angles in the state?
GyroCumulQuat+estimation.attitudeEulerAnglessuggest the former, unproven. - What each flag byte (this+0x1298..0x12aa, 0x1404) means, and which row count (1, 2, 3) maps to which sensor. The row weights are in FUN_0037c0f8; the flag-setting builders (FUN_00355600 and friends) are 330+ lines each, only partly read.
- Rest of FUN_0037a16c (pages 2 and 3 of 3 not read): update equations, joseph form, symmetrisation, and what happens after the “Kalman Gain computation failed” branch (skip update? reset?).
- Loop rate: the user-supplied “~90 Hz” conflicts with
adcFs 198.36 Hzand the “200 Hz” comments. Is ACQCON_colibry thread the 200 Hz estimation loop or a different, slower thread? Frame (0x9f0 bytes) content and rate need confirming from the producer side. visionIndicatorActiveandfusionVisionVzare false in common/estimator.cfg and not overridden in milosboard; that suggests vision speed observation is configured off by default, yet vision exports and gating logic exist. Confirm whether another cfg (video.cfg, states.cfg) or an API call turns them on (likely behavior-dependent, e.g. Hovering).- milosboard/barometer_compensation.csv does not exist in the rootfs although rcS expects
/etc/colibry/$(gprop ro.hardware)/barometer_compensation.csv. Check whatro.hardwareresolves to and whether the file is generated at first boot or by calibration. - libCondor to estimator coupling: how
Engine::getTranslationoutput (pixels) is converted to a velocity observation using altitude (ultrasound/baro), and whereRVisioncomes from. - Whether
Mykonos3::EstimatorState*FSM (Init, Error, RunningFlying) transitions are driven only by the immobility detector and baro-running condition seen at 0x004c593c. - I did not touch the physical drone or run any binary; Ghidra was used read-only (decompile, xrefs, strings, classes, callgraph).
Comments
Spotted a mistake, or tried this on your own Bebop? Leave a note. Comments are kept on this site's own server.