Knuckey RoverRMIT × Knuckey

The project

An autonomy layer for machines that already exist

Knuckey already builds the tractor and the seeder. This project tests, at prototype scale, whether a low-cost navigation, monitoring and safety architecture could drive them along a grower's mapped tramlines, and stop safely when something goes wrong.

The question

Everything links back to one sentence

How can a prototype autonomous CTF seeding vehicle integrate navigation, system-health monitoring and supervisory safety control to achieve accurate path tracking, while detecting hazards or critical faults and initiating a reliable safe-stop response?

Going straight to a full-size autonomous tractor would carry heavy technical, financial and safety risk, because localisation, obstacle response, fault monitoring and safe stopping all have to work together. Knuckey has suggested a horizon of three to five years for autonomous capability. This first year builds a scaled prototype and a simulation environment where the architecture can be tested and its weak points found before anyone commits to the full machine.

We treat simulated and prototype results as development evidence. A machine that drives itself is not, on that basis alone, a safe machine. We did not carry out a functional-safety lifecycle under ISO 25119 or ISO 13849, and we don't claim conformance with the ISO 18497 series.

Research questions

Four questions, and where each one stands

Each answer says what was measured, where, and what's still open. The Results page has the full numbers.

RQ1Accuracy

What line-following accuracy and repeatability does the navigation stack achieve on broadacre tramlines?

Met in simulation, not yet on the rover

Over 81 simulated passes, cross-track error was 18.3 mm RMS, or 3.8 mm once the first 20 m of each pass is left out. Neighbouring passes held their 12 m spacing to 5.7 mm RMS, and 15.8 mm with RTK-grade noise added. Sim

In the field the receiver repeated to 0.057 m lap to lap. Prototype 2, steered by a simpler route executor, sat 0.24–0.90 m RMS off the long rows. Field

RQ2Coverage

How does true coverage measured from logged vehicle paths compare with planner-level estimates?

Simulation only

A 12 m implement swept along the logged path covered 99.95% of the 670,680 m² target, leaving 312.8 m² untreated. 90.6% of the missed cells lie within 5 m of a pass start. Sim

Published planners report 97.29–100%, but they score the planned geometry. Ours includes tracking error and pass-entry behaviour, so the two aren't directly comparable. No implement was towed in the field.

RQ3Robustness

How severely does navigation degrade under brief GNSS dropouts, injected inertial disturbance and low-speed heading conditions?

Partly answered

Twice in the field the RTK corrections stopped. The receiver held its fixed solution for 600 s on ageing corrections, then dropped straight to a 3D fix of 1.6–2.1 m, and the rover kept driving. After a power interruption its heading reset to the magnetometer; switching on dual-antenna heading fixed it. Field

Vibration wasn't measured, because raw IMU data wasn't logged. The simulated dropout and IMU-noise studies are still being written up for the final report.

RQ4Safe stop

How reliably and rapidly does the layered stop architecture respond to communication failures and internal faults?

Fast where tested; reliability not yet measured

The physical E-stop pulled the main bus below 1 V in about 100 ms, and the rover stopped in 0.5 m from about 1 m/s. Rover Both faults that ended a run in autonomous mode were caught and latched by the rover's own monitoring. Field

The link-loss stop passed the software dry run but hasn't been tested on the moving rover. Most physical checks were single observations without logged timestamps.

Scope

What this year covers

In scope

Localisation
GNSS-RTK, IMU and wheel odometry fused through extended Kalman filters into one pose estimate.
Following
Consuming a supplied field plan, and measuring tracking and coverage against it.
Disturbance
GNSS loss, correction dropouts, vibration, heading error and low-speed turning.
Safety
Safe-stop behaviour, obstacle response and a layered emergency-stop design.

Out of scope

Product
A production-ready tractor or a certified agricultural safety system.
Mapping
Building the map on board. The rover takes the grower's existing plan.
Implement
Towing a seeder. Coverage uses a geometric 12 m footprint.
Avoidance
Planning around obstacles. Knuckey agreed to stop-and-wait for the prototype.
Accuracy against what?

Every tracking figure on this site is measured against the supplied waypoints. A perfectly followed waypoint can still miss the real tramline by whatever error the map carries. Published studies put RTK drone mapping at about 2–5 cm, the same size as the whole tolerance. Measuring against the ground needs an independent survey, which we didn't have.

Requirements

Targets and the evidence against them

Only customer and project requirements are judged met or not. Published results are context.

RequirementFromEvidenceStatus
Neighbouring passes within ±20–50 mmKnuckey site visitSim: 5.7 mm RMS (max 122.4 mm) ideal GNSS; 15.8 mm RMS (max 57.8 mm) RTK-grade. Rover: 0.24–0.90 m cross-track on the long rows.Sim as RMS only
Stop within 1 mCompletion Plan0.5 m at about 1 m/s. The stop path used wasn't recorded.Met as tested
Physical E-stop: motor current 0 A within 100 msCompletion PlanMain bus voltage below 1 V in about 100 ms on the bench. Current wasn't measured.Voltage proxy only
Safe state on an internal faultISO 18497-2Both run-ending faults in autonomous mode latched the E-stop: a bus collapse to 14.6 V and 1.01 s of stale voltage data.Shown, 2 events
Safe state on link loss, timed to stopping distanceISO 18497-20.5 s timeout, five missed 10 Hz frames, adds 0.5 m at 1 m/s. At the 1.67 m/s design speed it adds 0.84 m before braking starts. Dry run only.Software only
Defined response to lost positioningISO 18497-3The rover kept driving on a 3D fix after RTK was lost, twice.Not implemented
Obstacle stop inside 3 mProject settingStopped, waited and resumed indoors. In the field it took 500 mm to stop for a 1 m object, once, untimed.Not validated

Timeline

2026, from a shoebox-sized rover to a 50 kg one

Two semesters: Part A (OENG1167) in the first half of the year and Part B (OENG1168) in the second.

  1. Prototype 1 drivesA 3D-printed four-wheel skid steer with closed-loop motor control, wheel odometry and untethered PlayStation tele-op over its own Wi-Fi network.
  2. SLAM droppedBare soil offers too few features. The project switched to GNSS-RTK fused with IMU and wheel odometry.
  3. Site visit to Knuckey, WinchelseaThe ±20–50 mm requirement, John Deere compatibility, rocks as the main obstacle. What we heard.
  4. First real numbers0.22 m endpoint error over a 74 m walk. In Isaac Sim the dual-filter setup lifted waypoint completion from 28 to 32 of 34.
  5. Prototype 2 driveline and wheelsFour ODrive M8325s motors, 10:1 gearboxes, 260 mm wheels.
  6. 48 V power systemBattery mount fitted on 21 August; the drivetrain first ran with the wheels raised on 28 August; tele-op through the Pi worked by the end of the month.
  7. Prototype 2's first driveA gearbox support bracket added in mid-September made it run noticeably smoother.
  8. Stop testsThe remote E-stop stopped the wheels on both days. On 13 October the physical E-stop cut motor power, and every wheel held 36.7 rpm on the stand.
  9. Field test, Argyle SquareReceiver reference and walking test, then the rectangle and base-station row runs. Results.
  10. Out on grassThe photos and video on the home page.
  11. Final report and EnGeniusThis site goes with the team's booth poster.