Pathfinder system map

The six parts

Every term is defined in plain words in Terms.

Pathfinder is six named parts (owner decision D-NAME-1 "Domain names for the six parts", 11 Oct 2026). Every page, chart and test uses these names. "Physics" and "Fused" are no longer part names: they are the two modes of the Position Estimator, Pathfinder without ML and Pathfinder with ML.

Rust and Python identifiers, run IDs and file names keep their old spelling; renaming the code is a follow-up task.

Flowchart A · how the parts connect

flowchart LR
  SEN[Sensors
gyro, accelerometer,
barometer, compass, motors]
  VP[Velocity Predictor
velocity + uncertainty]
  PE[Position Estimator
weighted by uncertainty,
minus wind]
  NAV[Navigator
lat, lon, alt]
  AP[Autopilot
sees it as GPS]
  GF[GPS Fix]
  GIM[GPS Integrity Monitor
Healthy, Jammed, Spoofed]
  PA[Predictor Adapter]
  SEN --> VP
  SEN --> PE
  VP --> PE
  PE --> NAV
  NAV --> AP
  GF --> GIM
  GF --> NAV
  GIM -- switch --> NAV
  GF -- calibration phase only --> PA
  PA --> VP
  

Flowchart B · the two phases of a flight

stateDiagram-v2
  [*] --> Calibration
  Calibration --> Denied: GPS Fix lost or caught lying
  Denied --> Calibration: GPS Fix healthy again
  state Calibration {
    direction LR
    [*] --> Tuning
    Tuning: GPS Fix is the truth
    Tuning: Predictor Adapter retunes the Velocity Predictor
  }
  state Denied {
    direction LR
    [*] --> Flying
    Flying: GPS Integrity Monitor reports Jammed or Spoofed
    Flying: Adapter frozen
    Flying: Navigator sends the Position Estimator output as GPS
  }
  

1 · In flight: what runs on the drone

The flight controller (today the ArduPilot simulator, later a real autopilot) streams its sensors to a small companion computer. Pathfinder works out where the drone is and feeds that back. While GPS is trusted, ArduPilot flies on GPS and Pathfinder only watches and learns. When GPS is lost or caught lying, ArduPilot switches to Pathfinder's position and carries on with the mission.

flowchart LR
  subgraph FC[Flight controller - ArduPilot]
    S[Sensors
gyro + accelerometer 200 Hz
barometer, compass
motor outputs, GPS]
    EKF[ArduPilot EKF3
position used for flying]
    NAV[Autopilot
flies the waypoints]
  end
  subgraph CC[Companion computer - pathfinderd]
    IN[Read MAVLink
sensor messages]
    PROF[Vehicle profile
frame, motors, IMUs]
    subgraph CORE[Pathfinder core - Rust]
      F[Position Estimator
integrates the IMU]
      DRAG[Drag model
speed from tilt and air drag]
      ML[Velocity Predictor
air speed from IMU + motors]
      AD[Predictor Adapter
tunes the predictor to this drone
while GPS is good]
      GI[GPS Integrity Monitor
jamming and spoofing]
    end
    OUT[Send position
20 times a second]
  end
  S --> IN --> F
  IN --> PROF
  DRAG --> F
  ML --> F
  AD --> ML
  IN --> GI
  F --> GI
  GI -- GPS trusted --> AD
  F --> OUT --> EKF
  GI -- GPS bad: switch source --> EKF
  EKF --> NAV
  

Measured in the simulator only (D-10). The Velocity Predictor is currently the weak link: see "Where we are" below.

2 · Before flight: how the Velocity Predictor is made

We fly hundreds of simulated flights with varied wind, sensor errors and motor wear, record everything, and train a small network to predict the drone's speed through the air. The trained file is copied onto the companion computer.

flowchart LR
  G[Flight generator
random routes, wind,
sensor bias, motor wear] --> SITL[ArduPilot simulator
quad and octa]
  SITL --> LOG[Flight logs
sensors + true position]
  LOG --> SPLIT[Split by flight
train / check / test]
  SPLIT --> TR[Train network
PyTorch on the RTX 4080]
  TR --> ONNX[Model file
velocity-long.onnx]
  ONNX --> CORE[Pathfinder core
runs it with tract]
  

3 · Proof: how we test it

Every claim is a locked acceptance test (AC-1 to AC-9). Each test flies held-out scenarios the model never trained on, including a hexacopter it has never seen, and compares Pathfinder against ArduPilot alone.

flowchart LR
  SC[Held-out scenarios
jam, spoof step, spoof ramp,
GPS returns, clean] --> RUN[Test harness
flies each scenario
several ways]
  RUN --> B1[ArduPilot alone]
  RUN --> B2[Position Estimator
mode Pathfinder without ML]
  RUN --> FULL[Navigator
mode Pathfinder with ML]
  B1 --> AC[Acceptance tests
pass or fail with numbers]
  B2 --> AC
  FULL --> AC
  B1 --> V[Flight viewer
one HTML page per scenario]
  B2 --> V
  FULL --> V
  

4 · The simulation setup: what runs where

Everything runs on one Windows PC. The simulated drone is the real ArduPilot flight software flying a physics model, the same tool drone makers use. A ground station shows it live as a pilot would see it. The viewer shows what a ground station cannot: where the drone really was, against where each system thought it was.

flowchart LR
  subgraph PC[Windows PC]
    subgraph WSL[WSL - Linux inside Windows]
      LAB[Scenario lab
sim/lab.sh
pick wind, jam time,
spoof size, drone type]
      SITL[ArduPilot SITL
real autopilot code
+ physics model]
      PFD[pathfinderd
Pathfinder on the
companion computer]
      LOGS[Flight records
true position + every estimate]
    end
    GCS[QGroundControl or
Mission Planner
live view, like a pilot]
    VIEW[Flight viewer
HTML page per scenario
truth vs estimates]
  end
  LAB -- starts flights, jams or spoofs GPS --> SITL
  SITL -- sensors --> PFD
  PFD -- position --> SITL
  SITL -- TCP port 5763 --> GCS
  SITL --> LOGS
  PFD --> LOGS
  LOGS --> VIEW
  

5 · How Pathfinder hands its position to the autopilot

There are two ways to plug in. Both keep the GPS Fix in charge while it is healthy and switch to Pathfinder the moment it is jammed or caught lying.

flowchart TB
  PF[Pathfinder
position + how sure it is]
  subgraph A[Mode A - today: navigation feed]
    OD[ODOMETRY message
position, speed, confidence]
    EK[ArduPilot EKF3
external navigation input]
    SW[Source switch
GPS to Pathfinder on jamming]
  end
  subgraph B[Mode B - flown once: GPS replacement]
    GI[GPS_INPUT message
latitude, longitude, altitude,
speed, accuracy]
    G2[Autopilot sees a
second GPS receiver]
    SEL[GPS selection
prefers the healthy receiver]
  end
  PF --> OD --> EK --> SW
  PF --> GI --> G2 --> SEL
  SW --> FLY[Autopilot keeps
flying the mission]
  SEL --> FLY
  

Owner requirement R-3, 11 Oct 2026: the product must be able to feed the autopilot as a replacement for the GPS coordinates. Mode B: pathfinderd --output gps-input sends GPS_INPUT as the autopilot's second GPS (ArduPilot GPS2_TYPE 14) and makes it primary when the GPS Fix is untrusted (D-65). Mode A stays the default until Mode B passes the full test set.

For the presentation

• The problem: jam the GPS and a normal drone either lands or wanders off. In our simulator, ArduPilot alone is about 2 km off after two minutes.
• What Pathfinder is: a small program on a companion computer that works out position from the motion sensors every drone already has. No camera, no radio, no map.
• How it plugs in: it hands the autopilot its position as a navigation feed, or as an ordinary GPS signal, so any MAVLink drone can use it without changes.
• How we prove it: real autopilot software in simulation, scenarios the system never trained on, locked pass/fail tests, and a viewer that shows the true path against every estimate.
• Honest status: simulator numbers only; real flight data is the next phase.

Where we are, in plain words

ArduPilot-alone figures are medians over 8 held-out quad and octa flights; the Pathfinder rows are one flight (heldout-jam-1020). All simulator numbers.

The Position Estimator alone already works well. Adding the Velocity Predictor makes it worse, because the predictor's speed estimates are much less accurate than the estimator's own, and the estimator still trusts them too much. The next three steps:

1. Stop the current test run. It is testing the Navigator in Pathfinder with ML mode, which we already know is flawed, and would take hours to fail the remaining tests the same way.
2. Fly one scenario both ways and compare. One flight in Pathfinder without ML mode, one in Pathfinder with ML mode, side by side in the viewer, to confirm the Velocity Predictor is what drags it down.
3. Only listen to the predictor when it agrees. If Pathfinder without ML wins, the Position Estimator uses the Velocity Predictor only when it is close to what the estimator already says, and ignores it otherwise. Then rerun all the tests.