The archive · AI & Models · Product decision · 2026
Hebbian Robotics bets open, versioned data pipelines unblock robotics
YC S26 duo open-sources HFlow, an SDK that turns messy robot recordings into versioned, quality-checked datasets — Launch HN Aug 31, 2026.
Hebbian Robotics
What the business is
Hebbian Robotics builds HFlow, an open-source (Apache-2.0) SDK for robotics data pipelines: teams write plain-Python transforms, checks, labels and enrichments that HFlow runs as Airflow 3 DAGs, storing MCAP episodes with provenance plus a queryable DuckDB/Parquet catalog.
How it started
Brandon Ong had trained embodied-AI models for two-arm industrial cleaning robots and researched low-resource languages at Stanford; Kingston Kuan scaled systems at Jane Street and built Verkada's large-scale video platform. Both kept seeing robotics teams assemble fragile one-off scripts — one transcodes video, another checks timestamps, another labels — that quietly corrupt training data as corpora grow, and data vendors selling unverified egocentric or robot recordings.
What happened
The pair founded Hebbian Robotics in San Francisco during YC's Summer 2026 batch (founded 2026, two-person team) and released HFlow 0.2.x on PyPI through August 2026. The SDK processes synchronized video, state, actions and metadata in MCAP files, records provenance and version-stamped quality measurements in a Parquet catalog, and renders pipelines as visible Airflow graphs; the Launch HN on 2026-08-31 positioned it against teams rebuilding the same pipeline infrastructure themselves. The YC launch page pitches APIs that let data providers prove quality to buyers, with the hosted multi-tenant product still on a waitlist.
No ending yet — it is still running.
Background
Hebbian Robotics is a San Francisco startup founded in 2026 by Brandon Ong and Kingston Kuan during Y Combinator's Summer 2026 batch. Its product, HFlow, is an open-source SDK for robotics data pipelines: teams write transforms, quality checks, labels and enrichments as plain Python functions, and HFlow runs them in-process for development or as generated Airflow 3 DAGs for scheduled corpus processing.
The problem the founders attack is structural: robotics training corpora mix synchronized video, joint states, actions, timestamps and metadata from many recording systems, and teams usually start with scattered scripts that stop being auditable as data grows. Frozen cameras, drifting streams, missing topics and duplicate recordings slip into training data quietly. HFlow makes every processed episode carry its provenance, records quality measurements and version stamps in a Parquet catalog queried with DuckDB, and emits a version-pinned manifest instead of a hand-assembled folder.
HFlow deliberately picks standard boundaries: MCAP for recordings, Airflow for execution, Parquet and DuckDB for curation. The design principle is that existing robotics code should plug in through small adapters rather than being rewritten, and that the project is free under Apache-2.0 with no robot hardware required to contribute, which is how a two-person team hopes to become the default data layer for the field.
Public signals so far: the Launch HN on 2026-08-31 drew 46 points and 12 comments, the GitHub repository showed 126 forks in early September 2026, and RobotToday (from roboactu.fr) reported the 0.2.4 release with LeRobot v3 import on 2026-09-01. The software is pre-v1, the hosted multi-tenant control plane is not built, and the monetization plan — managed workspaces and enterprise support — is untested.
What has to be true
- Data quality is the bottleneck robotics teams hit first: a model cannot compensate for desynchronized, incomplete or irreproducible recordings, so QC belongs before training, not after failures.
- Standard formats beat another proprietary stack: MCAP, Airflow, Parquet and DuckDB are already trusted in the ecosystem, which lowers adoption cost for teams of any size.
- Open source is the wedge that fits a two-person team: Apache-2.0 code, contributions without robot hardware and visible provenance build trust with data buyers faster than salespeople could.
- Evidence over verdicts: checks record measurements with coverage and version stamps, so dataset policy stays with the consumer and every corpus stays reproducible.
What can be applied
Find the bottleneck customers feel first — data quality, not models — then open-source the infrastructure so standards compound, and sell hosting only after the format wins.
Aftermath
As of 2026-09-06 Hebbian Robotics is an active two-person YC Summer 2026 company in San Francisco. HFlow is pre-v1 with the core transform-to-manifest lifecycle working end to end, installable from PyPI, with 126 forks on GitHub and LeRobot v3 import added in 0.2.4. The Launch HN of 2026-08-31 drew 46 points and 12 comments, and the company is collecting names for a hosted platform waitlist while monetization through managed workspaces and enterprise support remains undeclared and untested.
Sources
- Launch HN: Hebbian Robotics (YC S26) – Build scalable robotics data pipelines
- Hebbian Robotics (YC S26): HFlow – open source SDK for scalable multimodal data pipelines in robotics and physical AI
- Hebbian Robotics Releases HFlow 0.2.4 for Robotic Data Quality Control
- Hebbian Robotics: APIs for robotics data teams to verify data quality for model training
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card