Overview

A data-operations console that records, synchronizes, and validates multi-stream robot episodes and manages versions and lineage.

Project Background & Purpose

The customer bundles six streams — arm joints, end-effectors, a head camera, and mocap — into 30 fps episodes for training datasets. Episodes with misaligned timestamps or missing frames used to reach the training queue and force retraining, with no history to roll back. The project measures alignment error against a sync-event jig, registers only episodes that pass a four-stage gate (schema, completeness, label quality, duplicates), and keeps every change as recoverable history from v38 to v41.

Components

• Six-stream recorder (arm joints ×7, end-effector pose, head RGB-D, 120 Hz mocap)
• Sync-event jig and PTP clock source (W-RC)
• Four-stage quality gate, version history, tiered storage (236 TB hot NVMe, 74 TB warm HDD, 2 TB quarantine)

Process Steps

1. 1Align the time base of all six streams with a sync event at session start
2. 2Check each saved episode in order: schema, completeness (under 0.5 % loss), label quality, duplicates (cosine 0.98)
3. 3Register passing episodes to a version and route rejects to the quarantine tier and rework queue