Hebbian Robotics is trying to fix the messy reality of robotics

Most robotics teams I talk to are essentially running a chaotic collection of "one-off" scripts to manage their data. You have one Python script to transcode video, another to check if the timestamps actually line up, a third to slap on some labels, and then a final bash script to copy everything into a training folder. It works fine when you have fifty recordings, but the second you hit five thousand, the whole thing collapses into a black box where nobody knows why a specific episode was excluded or if the dataset is even reproducible.

The YC S26 startup Hebbian Robotics is launching something called HFlow to address this exact bottleneck. It’s an SDK designed to turn raw, multimodal recordings—think synchronized video, joint states, actions, and sensor metadata—into standardized, quality-checked episodes and queryable manifests.

How the HFlow workflow actually functions

Instead of reinventing the wheel, HFlow seems to act as an orchestration layer for the tools that are already industry standards. Here is the technical breakdown of how they are approaching the problem:

• The Core Unit: They use MCAP (the open container format by Foxglove) as the primary input. This is smart because it functions like a ROS bag but is much better at keeping video, robot state, and sensor streams synchronized in a single file.
• Development vs. Production: During the R&D phase, you write your data processing steps as plain Python functions. You can run these in-process for quick testing. When you're ready to scale, HFlow packages those same registered steps as Airflow 3 DAGs. This allows for proper logging, retries, and task inspection in a production environment.
• The Cataloging Layer: This is where it gets interesting. Instead of constantly re-opening massive recording files to see what's inside, HFlow writes all measurements, metadata, and version stamps into an append-only Parquet catalog. You can then use DuckDB SQL to query your entire dataset and generate a version-pinned manifest instantly.

Solving the "Garbage In, Garbage Out" problem

The biggest headache in embodied AI is quality control. You don't want "impossible" joint movements or frozen video frames polluting your training set. Hebbian’s approach to quality checks is to treat them as reusable evidence rather than hard-coded rules.

They distinguish between two types of checks:
1. Deterministic checks: These catch the obvious stuff like black frames, timestamp drift, or missing topics without needing a heavy model.
2. Model-based checks: For more nuanced data, you can integrate VLMs (Vision Language Models) or tools like MediaPipe to detect specific hand poses or environmental contexts.

The key distinction is that HFlow doesn't just delete "bad" data. It quarantines it and records the evidence. This separates the policy (what you decided was bad) from the data (what actually happened), which is crucial when you realize your "bad" data was actually a goldmine for a different type of training task.

It’s a practical approach to a massive scaling problem. If you are building an AI workflow for physical robots, managing the data lineage is probably going to be more difficult than training the actual models.

https://github.com/Hebbian-Robotics/hflow