CROSS is our most robust mapping system yet. Darkness, moving objects, low-texture scenes and abrupt turns are the kinds of challenges that make mapping in real environments difficult. Our priority is a map the robot can keep using as those conditions change.
Watch the difficult stretches
This 59-second video follows four cases: a night run, fast motion with sharp turns in a corridor, night-time glare, and a corridor with the lights switched off. The camera view sits beside the full-session trajectories and mapping-error comparisons, so the conditions and the resulting map can be inspected together.
In the dark night sequence shown, CROSS stereo with images only completes the sequence while several feature-tracking baselines lose tracking. In the fast corridor case, the highlighted CROSS configuration uses wheel odometry. These are different sensor configurations, labelled in the video; their results should be read with those inputs in mind.
Tracking continuity and map quality both matter
A system can keep producing poses while its map drifts. The video therefore distinguishes lost tracking from whole-session trajectory error. It also includes systems that continue predicting poses in the dark, rather than treating every baseline as a tracking failure.
The clip directly illustrates darkness, glare, fast motion and sharp turns. Moving objects and low-texture scenes are broader mapping challenges we care about, rather than separate controlled tests demonstrated in this cut. The video reflects the benchmark snapshot used for its October 23, 2026 render; results vary by sequence and sensor configuration.