AI-generated illustration for Galileo Whispers. Conceptual rover wheel and terrain; not a named rover, measured slip, or a documented drive.
A Mars rover’s wheels can rotate through the equivalent of a long drive while the rover advances a shorter distance. Wheel encoders measure rotation; converting that rotation into travel assumes enough traction for the wheel to roll as expected. Visual odometry checks the motion of terrain features in successive images, providing another way to estimate how the vehicle actually moved.
The disagreement is useful information. It can reveal slip before a navigation system mistakes spinning wheels for progress.
What an encoder measures
For an ideal wheel rolling without slip on a simple flat path, multiplying its rotations by its circumference gives distance traveled. A rover’s software must also account for steering and the vehicle’s geometry. Those calculations still rely on assumptions about contact with the ground.
Loose soil can deform beneath a turning wheel. A wheel can dig in, push material backward, or lose grip on a slope. Its encoder may work perfectly while the conversion from rotation to motion becomes wrong.
Consider a hypothetical straight drive on level ground. The wheel-based estimate is 2.0 meters, but a separate reliable terrain-based estimate gives only 1.4 meters of advance. The difference is 0.6 meter. Defining slip for this example as that difference divided by the wheel estimate gives 30%. That definition and the straight-drive assumptions matter; they should not be silently transferred to every maneuver or slip convention.
A turn in place provides another caution. Wheels can rotate substantially while the rover’s center has little net displacement. That is not automatically loss of traction: the planned motion itself may be rotation rather than translation.
The ground supplies a second reference
Visual odometry identifies recognizable terrain features in camera images taken before and after a movement. With stereo cameras, features can be located in three dimensions. Their changed positions relative to the cameras constrain the rover’s translation and rotation.
A simplified example is a cluster of small rocks that appears closer and farther to one side after a short drive. One feature alone is ambiguous. Multiple features distributed through the scene help the software find a motion consistent with the group and reject mismatches.
Maimone, Cheng, and Matthies’s 2007 report on Spirit and Opportunity describes stereo visual odometry that estimates a six-degree-of-freedom pose change. That means three components of position and three of orientation, not simply a replacement number for wheel distance.
The images do not need to identify a named landmark in an orbital map for every step. Local tracking can estimate movement between observations without independently establishing the rover’s absolute planetary coordinates each time.
How the disagreement changes a drive
In a wheel-only estimate, the system might conclude that a desired advance has already been completed. If visual odometry instead finds substantial slip, it can update the estimated pose or contribute to a decision to stop or adjust the drive.
That matters near a target. A shortfall of several tens of centimeters can affect whether an instrument is in reach or whether a route stays clear of a hazard. On a slope, an unexpected sideways or backward component can matter even if the wheels rotated in the commanded direction.
JPL’s rover-navigation description separates visual odometry from hazard detection and path planning. Estimating where the rover moved, recognizing an obstacle, and choosing a new path are related but distinct jobs. A successful estimate does not prove that the terrain ahead is safe.
Images have failure modes too
Visual odometry needs trackable detail and a consistent interpretation of the scene. A largely featureless patch can leave too few points. Repeated patterns, substantial lighting changes, shadows, or an excessive movement between images can complicate matching.
In the MER system described in the 2007 report, internal checks could reject a visual estimate when tracking or convergence failed. Keeping an existing estimate in such a case is different from receiving an independent confirmation that it is correct. The software has gained no new visual correction simply because a camera exposure exists.
Wheel measurements, inertial sensing, image geometry, and consistency checks therefore work together. None is an infallible substitute for every other measurement. Historical success rates from one mission and period should not be treated as current universal performance specifications.
Distance traveled is not the whole pose
A rover can slip sideways, turn, climb, or partly slide back. Its accumulated path length, wheel-based travel, and straight-line displacement from the starting point can all differ without being contradictory quantities.
When reading a drive report, first ask what “distance” describes. Then ask how the estimate was obtained. Wheel rotation is a useful sensor measurement; actual vehicle motion is an inference that becomes stronger when independent terrain and orientation information agree with it.
