2026-11-14

Why Your Machine’s GNSS Heading Jumps at Low Speed

RTK fixed but heading jumps at a crawl? Why COG fails near zero speed, what single-antenna INS lacks, and how dual antennas and VIO help, plus a test plan.

A machine creeps up to a line, stops and backs up without steering. The receiver reports RTK fixed throughout. The heading on the screen wanders near the stop, then swings almost 180 degrees in reverse. The chassis never turns.

Start by checking which field the controller calls heading. Course over ground is the direction the machine is moving. Body heading is the direction the chassis is pointing. An inertial navigation system (INS) estimates body heading. On a machine that spends its day creeping in straight lines, the INS may have little information to correct that estimate as it drifts.

What RTK fixed tells you

RTK fixed means the receiver has fixed the carrier-phase ambiguities for its position solution. It says nothing about which way the machine points. With a single antenna, a receiver can derive course over ground from its horizontal velocity. At a crawl, even a small sideways error in the velocity estimate can throw off the angle. At a stop, course is undefined. Depending on the receiver and its configuration, the output can wander, go invalid or hold the last value.

When the machine reverses, its velocity points the other way, so course changes by about 180 degrees. The receiver can report RTK fixed through the whole maneuver. The Matrix guide to RTK correction services compares how network RTK and SSR corrections are delivered.

Check the interface before changing hardware. The course field in an NMEA RMC sentence reports direction of travel. If the steering loop needs chassis yaw, give it a body-heading estimate with its own validity and uncertainty. Trimble GNSS heading systems · NovAtel GPRMC message reference

Why a single-antenna INS can struggle at a crawl

The gyroscope measures angular rate, and the INS integrates that rate to track changes in yaw. Even a small gyro bias can cause heading to drift over time. See the Matrix guide to inertial navigation and drift for the IMU error sources behind that drift. At rest, gravity gives the INS a reference for roll and pitch. A turn about the vertical axis leaves that gravity reference unchanged, so gravity supplies no heading reference.

In a conventional single-antenna GNSS/INS, the filter can compare changes in GNSS velocity with accelerations measured in the machine frame. Starts, stops and turns can help it distinguish heading error from sensor bias. A long, straight crawl at nearly constant speed provides little information for that correction. A sharp turn at low speed can tell the filter more about heading than a fast, steady run.

Test how the system acquires heading and how well it holds it. Some systems initialize heading from course during a forward run with sufficient speed and suitable motion. A machine that powers up at rest and immediately reverses may never get that run. Even a system that aligns on a brisk open-sky drive can drift later if working motion gives the filter little to separate heading error from bias. NovAtel SPAN manual, kinematic alignment

What a second GNSS antenna changes

A dual-antenna receiver uses carrier-phase measurements to estimate the baseline between two antennas mounted on the machine. Once the receiver resolves that baseline, its azimuth gives a direct heading measurement at a stop, at a crawl and in reverse. The INS can use it as a yaw update while the baseline solution is valid. I’d start with dual-antenna GNSS if the machine needs a north-referenced heading at rest and can accommodate both antennas.

Use a rigid baseline with as much separation as the chassis and receiver allow. For a given transverse baseline error, a longer separation reduces its angular effect. Measure the antenna-to-IMU offsets and the angle between the baseline and the machine’s forward axis. A baseline across the machine can work if that offset is configured correctly. Check the antenna order too. Swapping primary and secondary can introduce a 180-degree error that looks a lot like a software problem.

Give both antennas a clear view where possible, and monitor the heading solution’s own status. One antenna can be shadowed by a boom or cab while the position still reports RTK fixed. Test on the machine with the boom or attachment in the positions it actually occupies on the job. Matrix’s GNSS guide for outdoor robotics covers compact GNSS/INS navigation for outdoor robots and off-highway machines. u-blox ZED-F9P moving-base application note · NovAtel SPAN dual-antenna configuration

Where visual-inertial odometry earns its place

Visual-inertial odometry (VIO) tracks features across camera frames and combines those observations with IMU measurements. As the machine crawls past a curb, a tree line or a building edge, those features shift in the image. Tracked features can constrain relative rotation even with little translation. That can limit yaw drift during slow maneuvers and GNSS gaps. Reliable translation needs usable parallax, and monocular VIO can struggle to estimate metric scale if the motion provides too little inertial excitation.

The camera needs features it can keep tracking. It can lose them over uniform grass, in dust or when the machine moves from bright sunlight into a dark bay. The Matrix guide to vision-based navigation covers it alongside SLAM and other alternative positioning methods.

At a stop, stable features can help constrain changes in orientation. VIO tracks changes in yaw relative to an initial heading and has no independent north reference, so an already aligned system can carry its heading through a GNSS gap. To establish an absolute heading at startup, use a valid GNSS baseline or another external heading reference.

Vision and dual-antenna GNSS can work together. The baseline anchors heading to north when satellite tracking supports it. Vision helps the estimator follow the machine’s motion when GNSS is degraded and the camera still has useful features. Hesch et al., camera–IMU observability analysis · Fixposition PBX-A1

Run the test at the machine’s working speed

  1. Cold-start at rest in open sky. Record how long it takes to get a valid chassis heading and whether the controller can tell when that heading is ready.
  2. Crawl straight, stop, then reverse without steering. Plot signed speed, GNSS course, baseline heading and fused yaw on one time axis. If the controller’s heading flips with course, trace the input all the way to the steering code.
  3. Turn slowly and repeat under representative obstruction. Include an in-place turn if the platform can do one. Run the route along a wall or under canopy, then repeat across the lighting and surface conditions the camera will see.

Time-align the logs and keep position fix, heading validity, measurement age, heading uncertainty and fusion state alongside the angles. Compare heading with an independent reference, such as a surveyed chassis alignment at rest or a separate reference system through the maneuver.

For each phase, report the 95th-percentile and worst heading errors, time without usable heading, and any intervals when the system reported a valid heading outside tolerance. When an aiding source returns, record how long heading takes to recover and how far it jumps. Use those results and the machine’s requirements to set the controller’s timeout and fallback behavior.

Use the machine’s path tolerance to put heading error in context. A one-degree pointing error projects to about nine centimetres sideways at a five-metre look-ahead. Actual tracking error also depends on the machine’s steering response, speed and control law.

For a machine that needs body heading before it moves, dual-antenna GNSS is a direct improvement over deriving direction from one antenna. For a machine that must keep tracking through low-dynamics maneuvers and GNSS interruptions, add vision where the route gives the camera enough to work with. The acceptance run should include the slowest maneuver the machine is expected to perform on its own.

If you’re evaluating hardware, send Matrix your minimum operating speed, usable antenna spacing and allowable heading error. We can help you define a test that matches the machine’s job. Request a Quote