04 / ML Engineering

HDF5 → Parquet / MP4

I built HDF5-to-Parquet-and-MP4 conversion to make robotics trajectories and videos usable in JEPA experiments.

R&D data preparationHDF5 → Parquet + MP4
STACKPython / HDF5 / Parquet / MP4
HDF5 pipeline with two outputs: Parquet tabular data and MP4 video sequences

01

The problem

HDF5 robotics datasets cannot be used directly by the target experimentation pipelines. JEPA experiments need independently usable video sequences and trajectories while preserving their correspondence within each episode.

02

My role

I designed and developed HDF5-to-Parquet-and-MP4 conversion to make robotics data usable in world-model experiments.

03

Constraints

  • Preserve relationships between tabular data and video sequences.
  • Produce formats usable across several experiments.
  • Handle the specifics of robotics and multimodal datasets.

04

Architecture & pipeline

Conceptual diagram
  1. Read source HDF5 data.
  2. Materialize tabular data and trajectories in Parquet.
  3. Export video sequences as MP4.
  4. Prepare these artifacts for training pipelines.

05

Decisions & iterations

Separating video and tabular data allows formats suited to each type of content.

Conversion is dedicated infrastructure, reusable beyond a single experimental protocol.

06

Results & limitations

The pipeline materializes tabular data in Parquet and video sequences in MP4 for JEPA / world-model experiments.

Video-to-trajectory correspondence remains a conversion constraint; the output schema and consistency checks still need documentation.

07

Next iterations

Document the output schema and consistency checks with conversion examples.

Other projects

Available for freelance projects

Need to improve an ML prototype?

Describe the errors you observe, the data available and your deployment constraints. We can define the next steps and the results to measure.

Contact meSaint-Étienne · Lyon · Remote · Available for freelance work · quotes on request