Launch HN: Hebbian Robotics (YC S26) – Build scalable robotics data pipelines

Aug 31, 2026 10:02 PM - 1 hour ago 3

Hebbian Robotics (YC S26)

Open root SDK for scalable multimodal information pipelines successful robotics and beingness AI

Y Combinator S26 Apache 2.0 license Join the Discord community

Hebbian Robotics (YC S26) is building HFlow, an unfastened root SDK for scalable multimodal information pipelines successful robotics and beingness AI. It makes information tooling and practices typically developed wrong ample robotics teams accessible to teams of any size.

We judge processing information is simply a awesome bottleneck successful robotics. A corpus tin combine video, state, actions, timestamps, and metadata from galore signaling systems. Teams often consciousness the problem first successful value control: determining whether cameras froze, streams drifted retired of sync, required topics disappeared, aliases copy recordings entered the corpus. As the corpus grows, fragmented scripts make it difficult to know what ran, audit the results, aliases reproduce a dataset.

Teams tin commencement pinch HFlow's built-in checks, constitute caller transformations, checks, labels, and enrichments, aliases link processing codification they already use. HFlow handles the orchestration, storage, versioning, and curation astir those steps.

HFlow stamps each processed section pinch its provenance, renders the pipeline as a graph, and records metadata and value grounds successful a queryable catalog. You tin trace really outputs were produced, show each stage, and investigate a corpus without loading the underlying recordings.

MCAP is HFlow's v1 input and output bound because it efficiently stores and serves synchronized video, state, action, and different time-series streams. That format request does not specify wherever the information comes from: human-worn cameras, teleoperated robots, autonomous policies, and different postulation systems can each provender the pipeline erstwhile their information is represented arsenic a supported MCAP episode.

Status: pre-v1, pinch the halfway lifecycle moving extremity to end. HFlow is fresh to effort locally. See what is implemented and open issues for existent specifications and remaining work.

Help turn the unfastened robotics community. Star the repository, stock it pinch your network, aliases contribute. Our extremity is an unfastened root organization wherever anyone tin participate successful building the early of robotics. No robot hardware is required to contribute.

HFlow's boundary

Input	Supported modular MCAP episodes directly; LeRobot Dataset v3 repositories done hflow import lerobot
Processing	Your Python transforms, checks, labels, and enrichments
Execution	In-process for development; generated Airflow 3 DAGs for scheduled runs
Durable output	Canonical MCAP episodes, provenance, artifacts, and a Parquet catalog
Curation	DuckDB SQL that writes a version-pinned manifest

Human and robot information move done a four-stage lifecycle:

collection --> ingestion ---------------> curation ------> delivery (landing (transform -> QC gross -> (SQL complete (curated MCAP + bucket) enrich, arsenic an section manifest; convert Airflow DAG) catalog) for training)

HFlow pipeline demo

Your processing codification stays yours. Transformations, value checks, labels, and enrichments are plain Python functions successful your ain environment. Existing codification plugs successful done mini adapters alternatively of being rewritten for a proprietary framework.
Episodes are MCAP, the instrumentality that ROS 2 records natively and Foxglove/Rerun unfastened directly, written pinch 2 tunings described successful Dyna's article: in-band H.264 pinch GOP magnitude matched to really the information is read, and topic-group chunking (camera streams and authorities streams ne'er stock a chunk, truthful a training sample costs 1 publication per group alternatively of 1 per topic).
Processed episodes transportation their provenance. The record itself records the schema, pipeline, and instrumentality versions that produced it, positive its root URI erstwhile available. Catalog records link measurements and outcomes to measurement versions, making it easier to trace a bad consequence backmost to its origin.
The pipeline is visible arsenic a graph. HFlow renders Airflow DAGs truthful you tin spot really stages link and show task status, logs, retries, and reruns.
Quality checks nutrient reusable evidence. Accessors extract the inputs existing processing codification expects (numpy arrays, MP4 paths, JPEG frames), and results onshore arsenic queryable measurements alternatively than hardcoded verdicts. Different datasets tin use different thresholds without processing the media again.
Query the corpus without loading the recordings. Metadata, value measurements, tags, type stamps, and artifact locations unrecorded successful the Parquet catalog. DuckDB tin reply corpus-wide questions and build manifests without opening the underlying MCAP files.

Open DuckDB's browser complete the catalog astatine immoderate time, including earlier the first run starts:

The open-source deployment is built to beryllium easy to own: tally 1 single-tenant workspace pinch the included Docker Compose runtime, aliases deploy its generated DAG bundle into an Airflow 3 situation you already operate. It has nary user accounts, RBAC, aliases multi-tenant power plane.

The information level is kept abstracted from relationship and control-plane concerns truthful the same motor tin beryllium scaled arsenic aggregate isolated workspaces (for example, one per squad aliases customer) down an outer power plane. That is the intended path to a early hosted version, but the hosted power level is not implemented successful this repository and is not a pre-v1 merchandise commitment. docs/HOSTING.md documents the data-plane statement that makes specified a power level an addition alternatively than a rearchitecture: the workspace unit, the seams a service drives (manifests, distant runtime addressing, credential injection), the spot model, and the existent limits.

Community and hosted interest

Hosted level waitlist: tell america astir your workflow.
Community Discord: join us for questions, feedback, and publication discussion.
Code of conduct: reappraisal our community standards and study concerns privately.

For reproducible bugs and scoped characteristic requests, use GitHub issues.

Install the SDK from PyPI pinch uv:

The Hebbian Robotics task starts astatine type 0.2.0. Earlier 0.1.x releases under the aforesaid PyPI sanction belonged to an unrelated, inactive task before the sanction was transferred.

To tally the repository's bundled quickstart:

git clone https://github.com/Hebbian-Robotics/hflow.git cd hflow uv sync --locked uv tally python examples/quickstart.py

The quickstart synthesizes a mini multimodal section pinch camera and state streams erstwhile nary input record is given, runs the pipeline in-process, and writes its outputs nether the gitignored data/ directory. It needs nary Docker or Airflow. To usage your ain recording:

uv tally python examples/quickstart.py path/to/episode.mcap

Use uv tally hflow --help to spot the CLI. When you are fresh to schedule the same pipeline, proceed pinch the runtime guide. Developers and contributors should commencement pinch CONTRIBUTING.md. Browse the examples catalog for the egocentric-corpus and OpenAI imagination paths.

To import a LeRobot Dataset v3 section into the aforesaid canonical MCAP boundary:

uv tally hflow import lerobot \ --repo lerobot/pusht --revision main \ --camera observation.image --episode-index 0 \ --output-dir ./data/lerobot_pusht

The importer resolves main to an immutable root perpetrate and records it as episode provenance. See the LeRobot import guide for the supported characteristic subset and a multi-camera example.

Get started successful six lines of code. This fuller illustration uses a robot teleoperation episode, but the aforesaid measurement interface applies to egocentric video and different physical-AI recordings.

import hflow from hflow.checks import camera_frame_stats from your_existing_qc import check_joint_smoothness # usage your existing checks app = hflow.App("kitchen-pipeline") # information root: $HFLOW_DATA_ROOT, hflow.toml, other ./data @app.check(version="1") def joint_smoothness(ep: hflow.Episode) -> hflow.CheckResult: joints = ep.channel("/joint_states").to_numpy() # our line: extract result = check_joint_smoothness(joints, rate_hz=100) # your line: unchanged return hflow.CheckResult(measurements=result) # our line: record @app.check(version="1", critical=True) def camera_blackout(ep: hflow.Episode) -> hflow.CheckResult: camera_topic = next(topic for topic in ep.cameras if "wrist_cam" in topic) evidence = camera_frame_stats(ep, cameras=[camera_topic]) black_frame_percent = evidence.measurements[f"{camera_topic}/black_frame_pct"] assert isinstance(black_frame_percent, float) return hflow.CheckResult( measurements={"black_pct": black_frame_percent}, verdict=black_frame_percent < 50.0, # percent; your threshold ) if __name__ == "__main__": app.test("episode_0001.mcap") # full pipeline, in-process, nary infra # Or telephone app.run() present to commencement the Compose runtime, past usage `hflow ingest`.

Every check, enrichment, and derived transmission declares a version. HFlow stores that worth precisely arsenic written: support it for behavior-preserving refactors, and bump it erstwhile aged and caller results should nary longer beryllium treated arsenic comparable.

Curation comes afterwards, via hflow.curate(data_root / "catalog", sql, output="manifest.parquet") or hflow curate "<sql>" connected the bid line, either measurement reporting coverage denominators alongside the manifest:

SELECT episode_id, uri FROM episodes WHERE task = 'fold_napkin' AND position = 'ok' AND black_pct < 1.0 -- percent, user-owned threshold AND pipeline_version = 'a41c9f27b3d8' -- pin 1 reprocessing generation

Democratize the architecture, defer the optimizations. Preserve the useful workflow and modular interfaces astatine mini scale, and explanation each production-scale system honestly arsenic implemented, simplified, deferred, aliases retired of scope.
Evidence, not verdicts. Checks grounds measurements pinch coverage; pass/fail argumentation belongs to the consumer, astatine curation time. Quality tags way episodes; they ne'er delete data.
Standard formats astatine each boundary. MCAP episodes, Parquet catalogs, Airflow DAGs. Our codification exists only wherever the format forces bridging aliases a pitfall is genuinely non-obvious.
Your codification stays your code. Existing transforms, checks, and enrichments plug successful done mini adapters alternatively of being rewritten.

Python ≥ 3.11
Docker (for the pipeline runtime; app.test() needs none), aliases bring your ain Airflow deployment (Astronomer, MWAA, Cloud Composer, self-managed)
The first hflow up downloads ~2 GB of instrumentality images and builds the task venv (one-time; app.test() needs nary of this)
Native s3://, gs://, and Azure information roots usage the optional bucket backend (uv sync --extra bucket); section paths do not import it
On Linux x86_64/aarch64, the first video cognition downloads a checksum-verified, pinned ffmpeg/ffprobe build into the personification cache. Set HFLOW_FFMPEG and HFLOW_FFPROBE to usage binaries you negociate instead.
Windows is supported via WSL2 (Airflow does not tally natively connected Windows)

Documentation home: commencement by task, past take a tutorial, how-to guide, reference, aliases explanation
Frequently asked questions: formats, infrastructure, scale, task scope, and existent merchandise status
How HFlow fits the robotics information stack: MCAP, Airflow, Foxglove, Rerun, DuckDB, entity storage, and training formats
Runnable examples: nonstop commands, prerequisites, expected output, and links to the applicable guides
Architecture and implementation status: the implemented, simplified, deferred, and out-of-scope matrix
Call OpenAI imagination from a step: a focused guideline linked to a complete executable pipeline
Contributing: improvement setup, validation commands, trial gates, and pull-request expectations
Security policy: supported versions and backstage vulnerability reporting

Dyna Robotics, Training Dyna-2 astatine million-hour scale, repeatably
MCAP specification and Python libraries (Foxglove)
foxglove.CompressedVideo schema: in-band H.264/H.265/VP9/AV1 video successful MCAP
Apache Airflow
DuckDB
Foxglove and Rerun
FFmpeg
Pareto, Hebbian Robotics' robotics information curation platform.

Thank you to each our contributors for making HFlow awesome! See CONTRIBUTING.md to subordinate our community.