RoboRun
RoboRun notes / physical AI data systems

Notes from the pipeline.

Technical writing on robot data validation, episode quality, labeling systems, and the compute between capture and training.

Publishing queue

The first field notes are in progress. This index will hold measured results, implementation details, and lessons from real data pipelines.

  1. 001

    The data bottleneck moved

    Why abundant capture makes detection, classification, quantification, and quality control the work that matters.

    In progress
  2. 002

    Quality gates for robot episodes

    What to measure before physical AI data enters behavior cloning, VLA training, reinforcement learning, or evaluation.

    In progress
  3. 003

    Decode once. Reuse the work.

    A systems view of sharing source encodings, frames, timestamps, and features across QA, labels, CV, training, and evals.

    In progress
What belongs here

Claims backed by measured runs.

Benchmarks, pipeline diagrams, failure analysis, dataset releases, and the engineering decisions behind RoboRun. No content treadmill—just work worth inspecting.