What changes between a measurement and understanding?
A research perspective · October 2026
A sensor does not observe “the world.” It records an interaction: light reflected from a surface, radio waves scattered by a leaf, or sound returning through foliage. The measurement contains information about a physical state, but that information is shaped by the way we observe.
Start with what we need to know.
For leaf wetness, the important quantity is water on the leaf surface. A visually detailed image is not automatically a good measurement of that quantity. A radar response is sensitive to physical properties, but interpreting it requires dealing with noise, geometry and the environment. Choosing a sensor is already choosing which evidence will be available.
Hydra combines camera and radar observations. Proteus uses a visual teacher to guide radar learning. Adonis moves from a wet/dry decision to a fine-grained wetness estimate. These are related but distinct answers to sensing questions. Their common thread is the relation between the observation, the representation and the quantity being estimated.
Keep the limits of the evidence.
A representation can discard nuisance variation. It can also discard something we will later need. A model that performs well under one set of conditions has not, by that fact alone, established what happens when the lighting changes, a sensor stops reporting, or the object differs from those seen in training.
That is why I want evaluation to expose the conditions of a result. What was observed? What served as ground truth? Which changes were tested? What remained outside the experiment? An error number becomes informative when these questions have answers.
Then ask what must persist.
Most of the published systems on this site answer a question about an observation or a short interval. A continuing understanding of a place asks for more: keeping track of what may have changed, carrying uncertainty when evidence is missing, and revising an estimate when new observations disagree.
The driving assessment preprint brings this question into a human setting. Behavior has context. The same observation can mean different things under different road conditions. The paper proposes design principles and research opportunities; a reliable monitoring system still needs to be built and evaluated.
Connect reliability to consequence.
My future agenda is to connect incomplete observations, persistent understanding and decisions. The goal is not to make every estimate actionable. It is to understand when the available evidence supports a decision, when another observation is needed, and when a person should remain responsible for the judgment.
That is the research program I want to build. The papers provide starting points. The questions about continuity and action remain questions.