Build a terrain model with known gaps

A navigation camera records light from a scene. A robot planning motion needs an estimate of shape and location. Stereo vision — open definition compares images from separated viewpoints and uses corresponding features to estimate depth. The camera separation, calibration and matching quality matter to the resulting geometry.

A picture that looks clear to a person can still be difficult for an algorithm. Uniform surfaces provide few distinctive matches; repeated patterns can create ambiguity. A dark region might be shadow rather than a hole. The model should preserve uncertainty rather than converting every pixel into a confident surface point.

Occlusion — open definition occurs when one feature blocks another. A boulder can hide a depression, and a ridge can conceal a steep slope behind it. Absence of a detected hazard in an occluded area is not evidence of safe ground. Unknown and traversable are different map states.

New viewpoints can reduce uncertainty, but acquiring them may itself require movement. A cautious plan chooses a position where additional sensing is feasible without relying on the very terrain that remains unobserved. This is a decision about the value and cost of another measurement.

A fictional route planner sees a short corridor with a shadowed section and a longer route observed from two useful viewpoints. If the mission requires verified clearance, the shorter path has an unresolved constraint. It may be worth taking another image before committing, or choosing the better-characterized route.

Distance is only one cost. Energy, slope, soil uncertainty, clearance and time to gather more information also matter. The lesson does not prescribe a rover’s actual hazard thresholds; those depend on its design and tested operating envelope.

Unknown terrain needs its own map state

As a robot advances, newly visible features can contradict the earlier terrain model. Updating the map is a normal part of exploration. Perseverance’s autonomous navigation illustrates the use of onboard images and terrain evaluation between instructions from Earth.

A map product should retain where the data came from, when they were obtained and how confident the interpretation is. When the model and new sensing disagree, slowing or stopping can preserve the opportunity to investigate. Confidence should follow evidence rather than the desire to keep moving.