# Vantar

> Vantar builds infrastructure for physical AI — retrieval and tooling for robots and embodied systems. Its flagship data product is Coldstack, a search layer over petabyte-scale robot logs.

## Coldstack — search over robot logs

Coldstack is a retrieval layer for robot data ("turbopuffer for MCAP"). Robotics teams keep their raw [MCAP](https://mcap.dev) logs in their own S3 bucket; Coldstack builds and hosts a compact index (about 1–3% of the raw size) on object storage, and exposes one query API that searches an entire fleet three ways at once — by metadata (robot, time, topics), by time-series signal (e.g. "torque above 5 for 2 seconds"), and by semantic meaning (image and text, e.g. "gripper slipping on a transparent object"). Matched moments export straight to LeRobot-compatible training datasets. Raw data never leaves the customer's bucket. Coldstack is early and recruiting design partners.

- [Coldstack overview](https://vantar.xyz/coldstack): what it is, who it is for, how the index and query API work, and the architecture (object-storage-native, index-only custody).
- [How robot-learning teams build training datasets from raw fleet logs](https://vantar.xyz/coldstack/building-training-datasets-from-robot-logs): why fleet logs are hard to search, the metadata + signal + semantic method for finding failures and edge cases across a fleet, and exporting them as LeRobot/RLDS datasets.
- [The MCAP file format, explained](https://vantar.xyz/coldstack/mcap-file-format): what an MCAP file is, what it is used for, how the format is structured, how to open and read one, and how to search across many MCAP files.
- [How to build a LeRobot dataset from your robot logs](https://vantar.xyz/coldstack/build-lerobot-dataset-from-robot-logs): what a LeRobot dataset contains, how to build one, and how to source good episodes from existing fleet logs instead of only fresh teleoperation.
- [Coldstack vs Foxglove](https://vantar.xyz/coldstack/vs-foxglove): the two solve different problems — Foxglove visualizes and debugs a recording; Coldstack searches a whole fleet to find the recordings worth opening and exports them as training data. They are complementary (Coldstack ships a Foxglove extension; both use MCAP).

### Coldstack by industry

Coldstack is tuned to the data each kind of fleet records. Segment guides:

- [Agricultural robotics](https://vantar.xyz/coldstack/for/agricultural-robotics): search camera logs from weeders, sprayers, and tractors; find model misfires and edge cases across a season.
- [Humanoid & embodied AI](https://vantar.xyz/coldstack/for/humanoid-robots): mine teleop and autonomy logs for the best demonstrations and informative failures; export as LeRobot datasets.
- [Autonomous delivery robots](https://vantar.xyz/coldstack/for/delivery-robots): find interventions and edge cases across a sidewalk/last-mile fleet.
- [Inspection & industrial robots](https://vantar.xyz/coldstack/for/inspection-robots): search years of quadruped/crawler/drone runs for a defect signature; keeps sensitive data in the customer's bucket.
- [Autonomous vehicles](https://vantar.xyz/coldstack/for/autonomous-vehicles): scenario and disengagement mining over petabyte camera/lidar/radar corpora.
- [Drones & aerial autonomy](https://vantar.xyz/coldstack/for/drones): find events and hard frames across a fleet's flights.

## Other products

- [Eve](https://vantar.xyz/eve): an open-source, server-native language experiment for typed conversations between distributed AI systems. Its compiler projects one global conversation graph into compatible endpoint plans and executes the same semantics over memory, TCP, authenticated QUIC, or Iroh. Machine-readable schemas include the [Eve Session Preface](https://vantar.xyz/eve/spec/eve-session-v0.schema.json) and [Eve Evidence](https://vantar.xyz/eve/spec/eve-evidence-v0.schema.json) contracts; the public [evidence protocol](https://github.com/Vantar-AI/eve/blob/main/rfcs/0004-evidence-protocol.md) pre-registers claims, baselines, and acceptance gates.
- [Nuro](https://vantar.xyz/nuro): open-source SDK for spiking neural networks — one API across neuron models, GPU today and neuromorphic silicon tomorrow.
- [Torq](https://vantar.xyz/torq): tooling in the Vantar physical-AI stack.
- [Cloud](https://vantar.xyz/cloud): Vantar's cloud platform.

## Contact

- Email: hello@vantar.xyz
- Coldstack design partners: robotics fleets curating training data from logs, or paying too much to search them, are the ideal fit.