HFlow is an SDK that converts multimodal recordings from robots and human operators—including synchronized video, joint states, actions, timestamps, and metadata—into standardized, quality-checked episodes and queryable dataset manifests. The pipeline exposes each step of conversion, checking, labeling, and enrichment as Python functions; these can be executed in-process during development or packaged as Airflow 3 DAGs for scheduled processing to inspect task states and retries.
Product LaunchesHebbian RoboticsHFlow
Hebbian Robotics Releases HFlow SDK to Convert Multimodal Robot Recordings into Structured Datasets
This article is a translation. Read the Japanese original
The input format is one MCAP file per episode, utilizing Foxglove's open container format. The output is a normalized MCAP containing in-band H.264 video and provenance describing the generation process, maintaining compatibility with Foxglove and Rerun.
Quality checks support both deterministic detection—such as black screens, frozen video, timestamp drift, and impossible joint movements—and model-based detection using VLMs and MediaPipe Hands. Each step is assigned a behavior version, and measurements, metadata, and artifact locations are recorded in an append-only Parquet catalog. Users can generate version-pinned manifests by querying with DuckDB SQL; episodes that fail critical checks are isolated, but the data is not deleted.
Founder Brandon stated that he faced the problem of robot data processing becoming a bottleneck for model improvement while training embodied AI models for dual-arm industrial cleaning robots, and Kingston reported facing similar issues while building high-throughput infrastructure at Jane Street.
Source: Launch HN: Hebbian Robotics (YC S26) – Build scalable robotics data pipelines (Hacker News Frontpage, 2026-09-01)