Coldstack for Autonomous delivery robots
Fleet log search for autonomous delivery robots
Your fleet drives public sidewalks and streets all day. Coldstack finds the interventions and edge cases across all of it — in one query.
Your data
Delivery robots run camera and lidar continuously across public sidewalks and streets. Every day adds terabytes per robot of normal driving, plus the rare moments that matter: interventions, near-misses, unusual obstacles, and the scenes your perception stack struggles with.
The problem
Edge cases are rare and scattered across the fleet
A confusing crosswalk or an unusual obstacle might appear once in thousands of hours. Finding every instance across the fleet is how you fix it — and it is exactly what manual review cannot do.
Interventions get logged but not mined
You know when a robot needed help, but pulling every similar situation out of the archive to retrain on is slow and manual.
Keeping it all searchable is expensive
Public-road data piles up fast. Hot storage to keep it queryable does not scale; cold storage cannot be searched.
What you can ask
Search your whole fleet in one query — by signal, image, and metadata.
- camera frames that look like a blocked sidewalk
- every teleop intervention in the downtown zone last month
- moments where the robot stopped for more than 10 seconds
- scenes with unusual obstacles near crosswalks
Matched moments export straight to a LeRobot-compatible dataset. Raw MCAP stays in your own bucket the whole time.
Questions
Can Coldstack find every instance of a specific edge case?
Yes. Describe the situation in plain words or give an example frame; Coldstack returns ranked matches across the entire fleet, so you can retrain on all of them, not just the one you remember.
Does raw footage leave our account?
No. Raw logs stay in your S3 bucket; Coldstack holds only the compact index and reads matched clips with short-lived signed URLs.
What log format do you support?
MCAP first — the de facto robot-log standard. ROS bag 1 and 2 are supported via conversion.