Coldstack for Inspection & industrial robots
Log search for inspection and industrial robots
Your robots build years of inspection runs. Coldstack finds a specific defect signature across all of them — without a full re-review.
Your data
Inspection robots — quadrupeds, wall-crawlers, and industrial platforms — record camera and sensor data on every run across facilities, and they accumulate years of it. The value is in comparing across the whole history: finding every place a defect pattern appears, or every run where a reading drifted.
The problem
Finding a defect across the whole history means re-reviewing everything
When a new failure mode is discovered, you want every prior run that shows the same signature. Without search, that is a manual pass over years of data.
Sensitive-site data cannot leave the customer's control
Inspection often covers sensitive infrastructure. Any tool that copies raw data out is a non-starter with security teams.
Cross-fleet patterns stay hidden
The same anomaly across multiple sites is a strong signal — but only if you can query the whole fleet at once.
What you can ask
Search your whole fleet in one query — by signal, image, and metadata.
- every run where a thermal or ultrasonic reading exceeded threshold
- camera frames that look like corrosion or a crack
- all inspections of asset 12 across the last two years
- runs where the robot flagged an anomaly and a human overrode it
Matched moments export straight to a LeRobot-compatible dataset. Raw MCAP stays in your own bucket the whole time.
Questions
Can we keep inspection data on our own infrastructure?
Yes — that is the design. Raw logs stay in your bucket with a scoped, read-only role; Coldstack custodies only the derived index. Nothing raw is copied out, which is what security teams need.
Can Coldstack search sensor readings, not just images?
Yes. You can query time-series signals (a reading crossing a threshold, for a duration) alongside image and metadata search, in a single call.
How much storage does the index add?
The index is roughly 1–3% of the raw log size, so making years of runs searchable does not mean paying to keep them all hot.