Data triage

Data triage is the practical business of deciding what recorded data to keep and what to throw away. A robot or an autonomous vehicle in operation produces an enormous volume of sensor and log data, easily more than can be stored on the device or shipped back over a network. Data triage is how teams cope with that flood, applying rules to prioritize the segments that are actually worth the storage and bandwidth, and letting the rest be discarded or overwritten. The name borrows from the medical sense of triage, sorting by importance when you cannot attend to everything.

The decisions often happen at the edge, on or near the device, because that is where the raw data first exists and where storage and connectivity are most constrained. Triage rules might keep everything around an incident or anomaly while retaining only a small sample of routine operation, or they might upload certain topics and drop others to save bandwidth. As physical AI systems scale to large fleets, this kind of selective retention becomes essential, and data triage has emerged as a recognized term for the discipline of doing it well.

Data triage applies rules and heuristics to the incoming stream of recorded data, scoring or flagging segments by how valuable they are likely to be. Common strategies keep data associated with incidents, anomalies, or rare events in full, while retaining only a fraction of routine operation, and they may selectively upload particular topics or time ranges rather than everything. On constrained devices, retention policies decide what to overwrite first when storage fills up. The aim throughout is to preserve the data most useful for debugging, analysis, and model improvement while shedding the redundant bulk that would otherwise overwhelm storage and networks.

Data triage matters because, at fleet scale, the cost and impracticality of keeping everything force hard choices, and those choices determine what a team can later learn from. For anyone operating physical AI systems, good triage is what ensures the rare, high-value moments survive while the redundant majority does not consume all the storage and bandwidth. Done poorly, it can silently discard exactly the edge cases and incidents that matter most, which is why it is increasingly treated as a discipline in its own right rather than an afterthought.

Because robots and vehicles generate far more data than can be affordably stored or transmitted, especially across large fleets. Triage prioritizes the most valuable segments so that limited storage and bandwidth are spent on data worth keeping.

It often happens at the edge, on or near the device, because that is where the raw data originates and where storage and network limits bite hardest. Deciding early what to keep avoids paying to move data that will just be discarded.

Typically segments around incidents, anomalies, and rare events, which are the most useful for debugging and improving models. Routine operation is often sampled sparingly, since keeping all of it adds cost without adding much new information.

Data curation, Scenario mining, Failure mining, Robotics observability, MCAP.