Mine Edge Cases From Petabytes of Fleet Recordings

How it works

1. Catalog your archive

   Connect S3, GCS, Azure, or Nebius. MCAP, ROS bags, parquet, and sensor files stay put while DataChain indexes metadata, lineage, and processing history.

2. Mine and label

   Find episodes by route, weather, or sensor version. Run labeling and embedding across 1 to 1000+ machines in your VPC, with the same Python SDK.

3. Version training mixes

   Save each mix as a named, immutable dataset version. The next VLA iteration reuses it, and every regression report points to the exact data behind the result.

From Fleet Archives to Training-Ready Data

Episode explorer

Find edge cases

nightrain

0042Low visibility00:18

0086Partial occlusion00:34

0117Sensor mismatch01:02

Event at 00:18

00:0000:45

night_rain_eval

Immutable training mix · v04

Retroactive mining

Mine edge cases without reruns

Every model iteration needs new training mixes. Query the catalog built by earlier labeling and extraction passes, instead of reprocessing petabytes of raw recordings.

• Filter, join, and similarity-search millions of episodes by route, weather, sensor version, label coverage, or anomaly.
• Find highway-rain events with CAN-bus anomalies as a catalog query, not a new processing job.
• Save each training mix as a named, immutable version for VLA training and exact regression references.
• Reuse previous processing results. Raw video, lidar, IMU, and CAN-bus bytes stay in storage.

Fleet catalog

EpisodesOne searchable layer

EPISODE / ROUTELABEL COVERAGE

0042 / highway

0086 / urban

0117 / highway

Source → every derived result

Brain

Query one shared fleet catalog

Turn each recording into a typed record. Fleet, route, weather, sensor versions, hardware revisions, and timestamps stay connected to every derived result through lineage.

• Know what exists: episode metadata, modality counts, and storage URIs.
• Track what’s done: 3D boxes, segmentation, tracking, calibration versions, and QA status.
• Find what matters: anomaly-flagged episodes, hand-labeled corner cases, and regression triggers.
• Add derived columns with each processing pass—one searchable catalog, not disconnected exports.

Cloud execution

autolabel_episodes.py

worker 01

worker 02

worker 03

Your VPC

Distributed compute

Compute the whole team can use

A laptop can’t autolabel 500K episodes overnight. Run labeling, embeddings, anomaly scans, and training-mix assembly from a notebook, script, or overnight agent.

• Attach CPU, GPU, and high-memory pools; scale from 1 to 1000+ machines with the same SDK.
• Workers start on demand and shut down when idle. No extra compute framework to manage.
• BYOC keeps compute in your AWS, GCP, Azure, or Nebius account, behind your VPC and existing IAM.
• Fleet bytes never leave your storage; DataChain orchestrates the work.

Agent workspace

From a question to a dataset.

Build a training mix from night-rain episodes.↵

Catalog and lineage loaded

Selection dispatched to 30 workers

night_rain_balanced

Version saved · v03

AI agents

Claude Code for fleet data

Give Claude Code, Cursor, and Codex the archive catalog, lineage, and clusters they need to mine edge cases and build balanced training mixes.

• Save each agent-mined dataset to the team registry with source code, parameters, episode IDs, and author.
• Dispatch a 30-machine query for night-rain events, then assemble a balanced training mix in one session.
• Generate typed datasets and distributed labeling pipelines; in the cited evaluation, vanilla agents materialized no reusable datasets.
• Measured cost-of-failure reduction: 2.7× overall and 9× on image work.