Hebbian Robotics, a YC S26 batch company, posted a Launch HN on August 31 for hflow, an open-source SDK for quality-controlling robot training data. The post scored 39 points with 10 comments—unremarkable. What caught attention was the repo's creation date: August 18, 2026.
In just 13 days the repo accumulated 211 stars, 191 commits, 149 issues (29 open), 182 PRs, 54 contributors, and 5 releases (latest v0.2.4). A companion repo, Pareto (Rust, created July 21), describes itself as the community edition of the company's self-hosted, self-operated data QC system—suggesting the SDK is a Trojan horse for a SaaS offering.
The problem it solves
Robot data collection channels have proliferated, but the quality of collected data is rarely discussed. A 2023 paper by Belkhale, Cui, and Sadigh (arXiv 2306.02437) studied robot data quality, and a later Voxel51 audit put numbers on it: of SayCan's 276,000 trajectories, only 12,000 (~4%) were usable; spot checks of open robot datasets found 20–40% of trajectories contaminated. Even the most basic metric—action smoothness—yielded an AUROC of just 0.244 when detection models ran on it, worse than a coin flip.
The hflow pipeline
hflow adds a factory-style QC pipeline in four stages:
1. Collection – MCAP audio/video + joint states + actions + timestamps 2. Ingestion – Python validation functions, generating Airflow 3 DAGs for scheduling 3. Curation – DuckDB queries over Parquet catalogs, version-pinned manifests 4. Delivery – clean datasets for VLA training
Every episode is written with provenance metadata (schema, pipeline, tool versions, source).
Two design decisions stand out:
- MCAP as the format bet – Foxglove-led and ROS 2's default log format, plus
hflow import lerobotfor LeRobot v3 data, aligning with the de facto standard. - "Evidence, not verdicts" – the pipeline only records measurements; threshold judgments are deferred to curation. QC rules and dataset versions are decoupled, so the same evidence can yield differently strict datasets. Anyone who has done data engineering knows how much overtime this separation saves.
- The README's headline performance figures—68% cost reduction at 43M episodes scale, 3.4×/2.9× speedups—are Dyna Robotics' data, with hflow merely noting the source and disclaiming affiliation.
hflow catalog uilaunches DuckDB's built-in browser UI, not a custom frontend.- The
hflowpackage name on PyPI was previously occupied by an unrelated 2020 project (by Hien Hoang); the README discloses the takeover honestly—a point in its favor. - 2026-07-21 – Pareto repo created, "community edition of our in-house QC system"
- 2026-08-18 – hflow repo created
- 2026-08-30 – v0.2.4 released, the 5th in 10 days
- 2026-08-31 – Launch HN; official YC page live
- Launch HN discussion
- hflow repository
- YC company page
- Robot data quality paper, arXiv 2306.02437
- Zhichai embodied AI daily, Aug 31 data observations
The founders
CTO Kingston Kuan lists Jane Street (million-messages-per-second systems) and Verkada (large-scale multimodal video) on his GitHub. Quant finance tolerates zero dirty data—one misaligned tick makes a backtest a lie. CEO Brandon Ong worked on cleaning robots at LionsBot and dropped out of a joint PhD program (Columbia/NTU). The YC page's "thousands of hours processed" claim comes from Kingston's own HN comment; the website's "million hours" vision should be treated as aspiration.
Caveats checked
Timeline
Why it matters
This year's embodied AI narrative has been all about models: bigger VLAs, better simulation. But the language-model era since Pythia proved the scaling game only wins if the data pipeline holds first. Robotics now has the world's most expensive data collection and the worst QC tooling—that scissors gap is a business. Goldman Sachs recently cut its humanoid robot average-price forecast to $21,300 by 2035; cheaper machines require falling marginal training costs per model generation, and data reuse is the one link nobody has seriously addressed.
hflow itself is still thin: more docs than code, delivery in a separate package, Windows only via WSL2. Thirteen days is too early to judge success or failure. But the position it occupies is more accurate than 90% of "yet another robotics foundation model" startups.
References: