Starling Uses Moving Landmarks to Ask Where It Is

What the demonstration adds

NASA’s August 17 Starling report describes FALCON, an experiment using onboard cameras, software and a catalog of known space objects to support navigation without relying on GPS for the demonstrated task. The agency presents this as progress in spacecraft autonomy. It does not mean every mission can now determine its position with any camera or that all external information has become unnecessary. The reported approach depends on an identifiable set of observations and reference information.

Its appeal is the reuse of a spacecraft’s view of surrounding objects. A camera can provide more than a picture for a human operator: with calibration and a suitable model, it becomes a source of geometric measurements that help estimate the vehicle’s state.

A star tracker usually answers an orientation question

Stars provide recognizable patterns that can help a spacecraft determine which way it is pointed. That orientation is different from knowing where the spacecraft is along an orbit. A distant star’s direction changes very little as a vehicle moves around Earth, so a star pattern alone does not provide every position component directly.

Nearby objects offer a different geometric opportunity. Their apparent directions depend more strongly on relative position, and their motion over time supplies additional information. The navigation problem is to connect those directions with known or estimated trajectories. This distinction is useful whenever a camera-based system is described broadly as knowing where it is: readers should ask whether the measurement establishes attitude, relative geometry or a complete orbital estimate.

A moving landmark needs a time-dependent address

A terrestrial landmark can often be treated as fixed for a navigation task. An orbiting reference moves substantially, so its catalog entry must describe more than a point on a static map. The observation time becomes part of identifying where the reference should be.

Catalog uncertainty matters because a navigation estimate inherits uncertainty from its references. If an object’s predicted position is imperfect, disagreement with the camera could come from the observer, the reference or both. The estimator needs enough information to distinguish or manage those possibilities. This makes catalog-based navigation an inference problem rather than a direct lookup of a spacecraft’s location from one recognized object.

Several observations can reveal what one cannot

A camera generally measures direction more directly than distance. One line of sight can correspond to many possible relative positions. Additional observations across time or toward different references help constrain the alternatives. The quality of the geometry influences how quickly and accurately the estimate becomes useful.

This is why an impressive single image does not establish navigation performance. A demonstration should explain the observation sequence and the reference used to evaluate the resulting orbit. It should also identify how the estimator starts when its initial position is uncertain. Solving a small correction around a good initial estimate is a different task from recovering from a much less informative starting condition.

Identification is part of measurement quality

A bright object in a camera view must be associated with the correct catalog entry before it becomes a reliable reference. Incorrect association can create an apparently precise but misleading estimate. The system therefore needs a way to test whether a match is consistent with other evidence.

Timing, image calibration and the spacecraft’s own orientation all affect that test. Reflections, faint objects and variable brightness can complicate detection. A robust method should be able to reject uncertain observations rather than force every visible feature into the solution. This is a recurring principle in autonomous sensing: declining to use weak information can improve reliability more than maximizing the number of measurements at any cost.

Autonomy changes where work happens

An onboard estimator can reduce dependence on a ground team for some navigation decisions. It still depends on software, reference information and a defined operational envelope. Autonomy moves part of the analysis onto the spacecraft; it does not remove the need to understand the analysis.

The benefits should be measured through useful performance under representative conditions. Computation time, power and response to missing observations all matter. A system that can maintain a good estimate during favorable viewing has established a capability, while operation through long gaps or unfamiliar conditions requires additional evidence. NASA’s report is most informative when read at the level of the demonstrated task rather than as a universal claim about GPS replacement.

A concrete example of uncertainty management

Suppose a spacecraft observes an object slightly away from its predicted direction. One explanation is that the spacecraft’s own orbit estimate needs correction. Another is that the object’s catalog trajectory is imperfect. A third is that the camera’s orientation or calibration is off. The observation becomes useful when additional evidence can constrain those alternatives, rather than when the software selects the first plausible explanation.

Observing several references and comparing their apparent motion over time can help. A common displacement across many objects might suggest a different problem from a discrepancy associated with one reference. The estimator can also compare a new observation with the state predicted from earlier measurements. These checks are a form of consistency reasoning that supports confidence in the result.

A practical navigation record should preserve which observations were accepted, which were rejected and how the uncertainty evolved. That allows a reviewer to understand whether an apparently accurate final orbit emerged from a stable sequence or from an accidental compensation of errors. For a vehicle making autonomous decisions, the history of confidence is as important as the final coordinates. It tells the rest of the system when the estimate is strong enough for a particular action.

How to evaluate future optical-navigation claims

Ask which quantities were estimated, which references were used and how accuracy was checked independently. Determine whether the reported result came from onboard processing during flight or from later analysis on Earth. Both can be valuable, but they support different claims about autonomy.

It is also useful to examine failure recognition. If the system loses sufficient references, can it report reduced confidence and preserve a reasonable prediction until observations return? An estimator should communicate uncertainty in a way that downstream control can use. A confidently wrong state can be more troublesome than an explicitly uncertain one.

Starling’s demonstration shows how existing sensor views can support a new navigational function when paired with suitable computation. Its broader lesson is careful use of geometry: moving landmarks can help establish location, provided their motion, identity and uncertainty remain part of the problem. That is a concrete engineering achievement, with future applications requiring evidence in their own environments.