HFlow

Product Introduction

1. Definition: HFlow is an open-source Software Development Kit (SDK) and data pipeline orchestration framework specifically designed for robotics and physical AI applications. It falls under the technical categories of MLOps (Machine Learning Operations) for robotics, multimodal data processing, and data curation platforms.
2. Core Value Proposition: HFlow exists to democratize and streamline the creation of scalable, reproducible, and auditable multimodal data pipelines. It solves the critical bottleneck of data quality control, processing, and curation for robotics teams by providing a structured framework that handles orchestration, storage, versioning, and provenance, allowing engineers and researchers to focus on their core data logic.

Main Features

1. Multimodal Data Pipeline SDK: HFlow provides a Python SDK that allows users to define custom data transformations, quality checks, labeling functions, and enrichment steps. These user-defined functions plug directly into HFlow's orchestration engine. The framework natively handles synchronized streams like video, robot state, actions, and timestamps, treating them as cohesive "episodes."
2. Provenance Tracking & Versioning: Every data processing step and its output are automatically stamped with provenance. This includes the pipeline definition, code versions, and source data identifiers. This creates a fully reproducible audit trail, making it possible to trace any dataset back to the exact conditions and code that generated it, which is essential for debugging and scientific rigor.
3. MCAP-Centric Storage & Parquet Catalog: HFlow uses the MCAP file format as its primary input and output boundary, optimized for efficient storage and retrieval of time-series multimodal data. Processed episodes are stored as canonical MCAP files. All metadata, quality measurements, tags, and artifact locations are written to a queryable Parquet catalog, enabling corpus-wide analysis without loading the raw media files.
4. Quality Checks as Evidence: Instead of hard-coded pass/fail verdicts, HFlow's check system records quantitative measurements (evidence) like black_frame_pct or joint_jerk. The pass/fail policy is applied later during the curation phase using SQL queries. This separation allows the same processed data to be re-evaluated with different thresholds for different training datasets.
5. Airflow DAG Generation & Orchestration: For production, HFlow compiles user-defined pipelines into Apache Airflow 3 Directed Acyclic Graphs (DAGs). This provides robust scheduling, monitoring, logging, and dependency management. The pipeline becomes visible as a graph, offering operational transparency into task status, retries, and execution history.

Problems Solved

1. Pain Point: Fragmented and unreproducible data processing scripts. Robotics teams often struggle with "script sprawl" – a collection of one-off scripts for data validation, filtering, and transformation. This makes it impossible to reliably know what processing was applied to a dataset, audit results, or reproduce a dataset for model training.
2. Target Audience: Robotics Engineers, ML Engineers in Robotics, Research Scientists in Physical AI, and DevOps/MLOps engineers supporting robotics teams. It is particularly valuable for teams scaling their data operations beyond a few researchers, including startups in the Y Combinator ecosystem and established companies building autonomous systems.
3. Use Cases:
   • Egocentric Video Corpus Curation: Processing hours of human-worn camera footage to filter out corrupted segments, extract relevant clips, and tag them with activity labels for imitation learning.
   • Robot Teleoperation Data QC: Automatically checking thousands of teleoperation recordings for issues like camera freeze, sensor dropout, or unnatural joint movements before using them for policy training.
   • Large-Scale Dataset Reproducibility: Ensuring that a dataset used to train a model six months ago can be recreated identically to benchmark new model iterations, including all preprocessing and filtering steps.
   • Multimodal Sensor Fusion Pipelines: Building a reproducible pipeline that synchronizes LiDAR point clouds, camera images, and IMU data, runs calibration checks, and packages them into training-ready examples.

Unique Advantages

1. Differentiation: Unlike monolithic, closed-source MLOps platforms, HFlow is open-source and focused exclusively on the unique challenges of physical and multimodal data. Compared to generic workflow orchestrators (e.g., plain Airflow, Prefect), HFlow provides built-in abstractions for robotics data (episodes, channels, cameras) and a strong opinion on using MCAP and evidence-based checks, reducing boilerplate code.
2. Key Innovation: The "evidence, not verdicts" architecture for data quality. By decoupling the computation of quality metrics from the pass/fail decision, HFlow enables flexible, SQL-driven curation long after processing. This allows teams to experiment with different quality thresholds for different training runs without reprocessing terabytes of data, a significant efficiency gain.

Frequently Asked Questions (FAQ)

1. What is HFlow used for in robotics? HFlow is used to build, manage, and scale data pipelines for robotics and physical AI. It helps teams ingest raw sensor data (like video and robot state), run automated quality checks, apply transformations and labels, and curate high-quality datasets for training machine learning models, all while maintaining full reproducibility and provenance.
2. How does HFlow handle data versioning and reproducibility? HFlow automatically tracks the version of every user-defined function (check, transform) and the pipeline itself. This information, along with the source data ID, is stamped into the output metadata and catalog. To reproduce a dataset, you specify the pipeline version and curation SQL query; HFlow's architecture ensures the same data and code versions are used.
3. What is MCAP and why does HFlow use it? MCAP is an open-source container format optimized for robotics and automotive time-series data, supporting mixed binary and structured data like video streams and sensor messages. HFlow uses it because it's efficient, self-contained, and widely supported by tools like ROS 2, Foxglove, and Rerun, making processed data immediately usable in the robotics ecosystem.
4. Can I use my existing Python data processing code with HFlow? Yes. A core tenet of HFlow is "your code stays your code." Existing data validation, transformation, or labeling scripts can be wrapped in small adapter functions to fit HFlow's check, transform, or enrich decorators, allowing them to integrate into the managed pipeline without a full rewrite.
5. Is HFlow a hosted service or self-managed software? Currently, HFlow is primarily a self-managed, open-source SDK and runtime. You can run it locally or deploy it on your own infrastructure using Docker Compose or an existing Airflow cluster. The creators (Hebbian Robotics) have designed it with a separable data plane to enable a future hosted control plane, but the hosted version is not yet available.