RTK Meets IMU: Why Can a Continuous Track Still Be Wrong?
On this page
TL;DR
An RTK + IMU receiver can keep reporting position through a satellite outage while accumulating meters of error. Inertial measurements carry motion forward between GNSS updates; sensor noise and bias accumulate during integration. Reliable GNSS observations correct the estimate, while wheel-speed aiding adds a constraint on forward motion.
From RTK and OSR/SSR to PPK, the approaches covered in this series depend on satellite observations. Buildings, tunnels and reflected signals can make those observations unreliable or unavailable. A receiver that works well in open sky still needs a way to handle those gaps.
Vehicle navigation often combines GNSS with cameras, LiDAR or millimeter-wave radar. An inertial measurement unit (IMU) offers a compact, relatively inexpensive source of motion data. It senses acceleration and rotation on the vehicle itself, so its measurements continue through darkness, covered parking and tunnels.
What Does a Six-Axis IMU Actually Measure?
Six axes, two kinds of measurement
A six-axis IMU combines a three-axis accelerometer with a three-axis gyroscope (Figure 1). The accelerometer measures along X, Y and Z, responding to acceleration, braking and orientation relative to gravity. The gyro measures angular rate: how fast the vehicle rotates about each axis, commonly in °/s. Angular acceleration describes how that rate changes.
A typical sample contains a timestamp, three accelerometer readings in m/s² and three angular-rate readings. These six channels are the starting point for calculating movement.
From measurements to movement
Integration adds up small changes over time. Integrating gyro measurements updates attitude: rotating about a fixed axis at 10°/s for 2 s changes orientation by 20°. During three-dimensional motion, the calculation accounts for changing axis directions and rotation order.
That attitude lets the system express accelerometer readings in navigation directions and account for gravity. Integrating the resulting motion acceleration updates velocity; integrating velocity updates position. Both calculations start from an existing navigation state.
From an IMU to inertial navigation
An IMU supplies the measurements. An attitude and heading reference system (AHRS) uses them to estimate orientation, while an inertial navigation system (INS) also estimates velocity and position. GNSS observations can then correct the INS as the vehicle moves.
Dead reckoning (DR) carries a known position forward using motion information, including IMU data or wheel odometry. In GNSS + IMU dead reckoning, inertial propagation carries the estimate between satellite updates. Receivers may report a DR or Estimated solution state; that reported state identifies the DR interval in the example below.
How Do RTK and IMU Work Together?
RTK provides a precise satellite-based position reference in clear sky. The IMU tracks motion between updates. A fused estimator combines them to follow position, velocity and attitude as the satellite view changes.
Carry motion through a short GNSS gap
When a vehicle enters a tunnel, the navigation system starts with its last position, velocity and attitude. New inertial measurements update that state through braking, acceleration and turns while the satellite view is blocked.
This short-term propagation helps vehicle tracking through underpasses, short tunnels and covered-parking entrances.
Keep track of orientation while moving
Attitude connects the sensor's moving axes to navigation directions. It also helps a mobile-mapping system place camera images or LiDAR measurements as the vehicle turns and tilts, including in clear sky.
Update the state between GNSS fixes
IMU measurements, GNSS position updates and fused navigation output serve different jobs and can run at different rates. Table 1 and Figure 2 use this example:
| Information | Rate | What arrives |
|---|---|---|
| IMU measurements | 100 Hz | New motion measurements every 10 ms |
| GNSS position updates | 1 Hz | An external position update every second |
| Navigation outputs | 20 Hz | A current state estimate every 50 ms |
In Figure 2, GNSS updates arrive once a second, while new IMU samples support a navigation output every 50 ms. Each output reports the current motion estimate. The GNSS solution rate is separately configurable and can also exceed 1 Hz; the three rates describe different parts of the receiver.
How Do GNSS and a Kalman Filter Keep IMU Drift in Check?
Integration accumulates measurement errors too
Accelerometers and gyroscopes contain random noise and slowly changing biases. Integration carries those errors into velocity, attitude and position along with the real motion. A persistent acceleration bias creates a growing speed error, then an increasing position error. A gyro bias can rotate the estimated travel direction; a tilt error also projects part of gravity into the horizontal acceleration estimate.
This explains why simply integrating faster is not enough. The system needs observations that help distinguish sensor error from actual movement.
Prediction, comparison and feedback
The Kalman filter, introduced by R. E. Kalman in 1960, provides a way to combine a prediction with new measurements. It tracks the estimated state and its uncertainty through a repeating cycle: predict the motion, compare with an observation, correct the estimate, then carry that correction into the next prediction.
Suppose inertial integration gradually overestimates forward speed, while GNSS velocity shows steady travel. That difference gives the filter evidence of an error. It adjusts the navigation state according to the uncertainties assigned to the prediction and measurement, and can estimate sensor biases that the available motion and observations make identifiable.
This feedback helps the estimate converge as reliable observations arrive. Inside a tunnel, wheel speed or vehicle-motion constraints can still help, but absolute position uncertainty grows without a position reference. Under an elevated road, reflected GNSS signals pose a different problem: a biased update can pull the fused track off course. The estimator therefore also needs to assess the quality of incoming observations.
How Does Wheel Speed Help the IMU Track Motion?
Constraining forward speed and distance
Wheel pulses, encoders or a suitable vehicle-speed source provide another observation of how fast the vehicle moves and how far it travels. That reduces reliance on acceleration integration alone. Gyro measurements describe turning while wheel speed helps constrain forward movement.
If inertial speed drifts upward while wheel speed stays steady, the estimator can use the difference to correct its speed estimate. It needs wheel pulses or a time-stamped vehicle-speed input to make that comparison.
Wheel speed alone supplies neither absolute position nor absolute heading. Correct distance projected in the wrong direction still produces a wrong track.
Vehicle motion constraints: NHC and ZUPT
Under suitable rolling conditions, the vehicle's reference point can be treated as having approximately zero sideways and vertical velocity in the vehicle frame. These non-holonomic constraints (NHC) do not prevent turning or climbing a slope; they are not zero vertical speed in a geographic frame.
When the system reliably detects a stop, a zero-velocity update (ZUPT) can add another constraint. Wheel speed is a measurement, NHC describes assumed motion, and ZUPT depends on correctly identifying rest. Slip, timing errors, incorrect travel direction or false stop detection can make aiding misleading.
The GT-01 example below shows how the receiver performed using GNSS + IMU without wheel-speed input.
How Much Inertial Drift Can Your Application Tolerate?
Small percentages can become meters
A low-cost IMU can make a short positioning gap easier to bridge. Whether the resulting position is useful depends on the job: recording which road a vehicle followed and controlling its position to a few centimeters allow very different errors.
A small percentage of distance can translate into meters over a long outage. These hypothetical endpoint-error ratios show the scale:
| Endpoint error / distance | 100 m | 500 m | 2 km |
|---|---|---|---|
| 0.1% (1‰) | 0.10 m | 0.50 m | 2 m |
| 0.5% (5‰) | 0.50 m | 2.50 m | 10 m |
| 1% (10‰) | 1 m | 5 m | 20 m |
At 0.1%, a 2 km passage ends 2 m off. That may still identify which road a vehicle followed, while exceeding the clearance available to a small robot. Use the ratio to translate a quoted result into distance, then compare that distance with the task.
IMU performance is part of that choice. Consumer and industrial MEMS devices, and higher-performance optical gyros, offer different noise, bias and cost trade-offs. A grade label is a starting point for comparing sensors; the position error budget belongs to the complete navigation system.
Time and recovery belong in the budget
Driving 1 km quickly differs from creeping and stopping over the same distance. Inertial errors depend on time, motion, initial state, environment and aiding. Some errors can change even while distance stops accumulating.
Set a maximum acceptable error before testing, then include the longest likely outage and the period after GNSS returns. Wheel-speed availability, turns and stops all belong in that test.
RTK + IMU in a Tunnel: What the Test Shows
Continuous 10 Hz output, meter-level error
In Kalmix's GT-01 test report, recorded on 8 June 2026, the GUIDE-K35E receiver maintained 10 Hz position output through a predefined tunnel scenario of about 2.45 km and 108 s, without wheel-speed input. Figure 3 follows the position error through the passage and into GNSS recovery.
Before reporting DR, the receiver passed through Fixed, Float and differential positioning states. Horizontal error was already about 0.24 m at the start of DR. Propagation inherited that starting estimate. The continuously reported DR interval lasted about 92 s over 2.17 km.
We compared the receiver and reference at matching timestamps. Horizontal error includes both along-track and sideways separation; signed vertical error is receiver height minus reference height. The results show the scale of the drift and subsequent correction:
| Observation | Horizontal error | Vertical error |
|---|---|---|
| DR peaks (at different times) | About 4.32 m | About 1.50 m absolute |
| Last reported DR point | About 2.32 m | About −0.34 m |
| First Fixed point during recovery | About 0.064 m | About +0.065 m |
In Figure 3A, horizontal error reached 4.32 m about 41 s into DR, then fell and fluctuated. The largest deviation occurred inside the gap rather than at its endpoint. For a vehicle working near a boundary, that peak can matter more than the final error.
What happened as GNSS returned
In Figure 3B, horizontal error was about 2.32 m at the last DR point. The first Float update brought it to about 0.27 m, followed by a brief rise to about 0.75 m. Fixed returned 2.2 s after that Float update, with horizontal and vertical errors both around 6 cm. The time axis follows these logged state changes, showing how the position estimate corrected as GNSS returned.
The continuing log stayed within 10 cm horizontally and vertically for just over a minute from the first Fixed point. For the vehicle using this stream, recovery includes both the renewed accuracy and the abrupt change from the drifted position.
For a tracking application, the continued output may be valuable. For lane assessment or control, compare the deviations and recovery jump with the allowed position error.
What to Look for in an RTK + IMU Track
The practical gain from inertial aiding is that the receiver can follow movement through a GNSS gap. For vehicle tracking, that can preserve a useful route history. For a machine operating near an obstacle or lane boundary, the decisive number is the largest deviation from the reference, including the correction as GNSS returns.
Key Takeaway
Evaluate RTK + IMU against the job the position must do: continuity through the gap, the largest deviation along the way, and the correction when GNSS returns. Set an acceptable error first, then use those three measures to decide whether the system fits.
Frequently Asked Questions
What is the difference between an IMU, an AHRS and an INS?
An IMU supplies three-axis accelerometer and gyroscope measurements. An AHRS uses motion measurements to estimate orientation, while an INS also calculates velocity and position from an initial state. Adding GNSS observations gives the navigation system a position and velocity reference for correcting accumulated inertial errors. The distinction is between measuring motion, estimating orientation and following a trajectory.
Can a Kalman filter remove IMU drift?
A Kalman filter corrects inertial drift by comparing its prediction with observations such as GNSS position, GNSS velocity or wheel speed. The discrepancy feeds back into the next estimate and can help identify sensor biases. During a GNSS outage, wheel speed and motion constraints can still limit some errors, while uncertainty in absolute position grows until a position reference returns.
Does a 20 Hz navigation output mean 20 RTK fixes per second?
A 20 Hz navigation output reports 20 state estimates per second, even when the RTK engine provides fewer fixes. The example here uses 100 Hz IMU samples and 1 Hz GNSS position updates to produce an estimate every 50 ms. Inertial propagation follows motion between the GNSS updates. Read these three rates separately in a receiver specification: IMU measurement rate, GNSS solution rate and navigation output rate.
How long can an IMU navigate without GNSS?
The usable duration ends when position error reaches the application's limit. In the GT-01 tunnel passage, the receiver reported DR for about 92 s over 2.17 km without wheel-speed input, and horizontal error peaked at 4.32 m. That gives a concrete scale to compare with your tolerance. Sensor performance, starting error, turns, stops and aiding determine how another passage develops.
Why does wheel speed not eliminate dead reckoning drift?
Wheel speed constrains forward speed and distance, while heading error can still project that distance onto the wrong path. As travel continues, a small direction error creates sideways separation from the true track. Wheel slip, tire-radius changes and timing errors also affect the measurement. Combining wheel speed with gyro measurements improves the motion estimate; GNSS supplies the absolute position reference for correcting accumulated drift.
Why can an RTK + IMU position jump when GNSS returns?
Returning GNSS observations correct a position estimate that has drifted during the outage. In GT-01, horizontal error changed from 2.32 m at the final DR point to 0.27 m at the first Float update. It reached 0.064 m when Fixed returned 2.2 s later. A controller consuming that position stream needs to handle the correction as well as the recovered accuracy.
Continue exploring
GNSS Errors Mapped
Understand which satellite-positioning errors remain before inertial aiding enters the picture.
LateralRTK Accuracy Trap
Read field errors and solution states together before trusting a headline accuracy figure.
ApplyRTK GPS for Robotics
Connect GNSS, IMU and odometry to the behavior an outdoor robot needs when positioning degrades.
