Hebbian Robotics Launches HFlow, an Open-Source SDK for Robotics Data Pipelines
Brandon and Kingston, founders of Hebbian Robotics (YC S26), have launched HFlow, an open-source SDK designed to standardize and quality-check robotics training data pipelines. The tool processes multimodal robot recordings — including synchronized video, joint states, actions, and timestamps — stored in the MCAP file format into structured, reproducible datasets. HFlow addresses common pipeline failures such as frozen cameras, timestamp drift, and duplicate recordings by storing deterministic quality-check evidence in an append-only Parquet catalog queryable via DuckDB SQL. For scaled processing, the SDK packages pipeline steps as Airflow 3 DAGs, allowing teams to monitor task status, logs, and reruns without rewriting existing infrastructure. The founders developed the tool after independently encountering robotics data bottlenecks — one while training embodied AI models for industrial robots, the other while building high-throughput infrastructure at Jane Street.
