Coldstack for Autonomous vehicles
Scenario and log search for autonomous vehicle fleets
Your fleet records camera, lidar, and radar all day. Coldstack mines the scenarios and disengagements out of all of it — as one query.
Your data
AV fleets produce some of the largest robot-data volumes anywhere — camera, lidar, and radar recorded continuously, plus a rich telemetry stack. The work is scenario mining: finding disengagements, rare events, and specific driving situations across petabytes, fast enough to feed training and evaluation.
The problem
Scenario mining across petabytes is slow and costly
Finding every instance of a cut-in, an unprotected turn, or a specific disengagement across the corpus is core to AV development — and it is expensive to do at scale.
Most teams build this in-house and it is a huge effort
AV data infrastructure is a major internal project. Smaller and mid-stage teams cannot always justify building and running it.
Cold data is unsearchable, hot data is unaffordable
The corpus is too big to keep hot; archived, it cannot be searched — so long-tail scenarios stay out of reach.
What you can ask
Search your whole fleet in one query — by signal, image, and metadata.
- every disengagement on an unprotected left turn
- scenes that look like a cyclist entering the lane
- moments with hard braking above a threshold near intersections
- all runs in heavy rain from the west-side fleet
Matched moments export straight to a LeRobot-compatible dataset. Raw MCAP stays in your own bucket the whole time.
Questions
How does Coldstack differ from an in-house scenario-mining pipeline?
It is the same capability — search the whole corpus by signal, scene, and metadata — without building and running the pipeline yourself. You point it at your bucket and query; the index lives on object storage.
Does our raw sensor data leave our account?
No. Raw logs stay in your bucket; Coldstack holds only the index and reads matched slices with short-lived signed URLs.
Can it handle petabyte-scale corpora?
Coldstack is built natively on object storage for exactly this — a bounded, cheap read per query instead of an always-on hot database. Fleet-scale latency is still being benchmarked with design partners.