Anchor

Anchor is HCS Labs' data refinery for humanoid robot learning: a pipeline that turns ordinary human video into provenance-verified, feasibility-checked, robot-ready training data.

Project page: https://hcslabs.github.io/anchor/

Intake → Trust → Gates → Enrich → Retarget → Package → Serve

What's in this repo

index.html      the project page (no build step — just open it)
assets/         diagrams, validation imagery, demo video clips
pipeline/       the pipeline SDK — open stages as working code, closed
                stages as interface stubs. Start at pipeline/README.md
third_party/    credits for the published / open-source work the
                reconstruction and intake stages build on

The SDK, in one breath

cd pipeline && pip install -e ".[ingest]"
anchor probe clip.mp4      # probed + content-addressed episode record
anchor gates clip.mp4      # open quality-gate cascade verdict
anchor stages              # the seven stages, open vs closed

Working code in the SDK today: scoped source connectors (licensed-API acquisition, local/partner drops), ffprobe intake with content addressing, the quality-gate cascade (open tier), CLIP-based prompt-driven person segmentation, and the Noise-as-Signal residual sampler — plus four runnable examples. See pipeline/README.md for the full module map and the honest open/closed table.

Noise as Signal

New on the page (§04.5): our research direction treating the residual between video-based pose estimation and mocap as a structured, task- conditioned noise distribution — and injecting it during policy training as a regularizer for sim-to-real transfer. Results are being validated internally; a public sampler skeleton lives in anchor.noise.

Validation

The reconstruction stage builds on VideoMimic (CoRL 2025); we re-ran its own evaluation protocol (SLOPER4D filtered subset) and report our measured numbers on the page — clearly separated from cited baselines. Everything downstream of reconstruction — acquisition, multi-robot retargeting, feature extraction — is our own work.

What's deliberately withheld

The Trust + Gates classification logic and the retargeting correspondence / shape-fitting method. Diagrams mark these with dashed borders and a lock; code marks them with interface stubs that raise NotImplementedError.

Contact

Questions about Anchor, research collaborations, licensing, or commercial partnerships are welcome.

Email: contact@hcslabs.ai

Ownership

© HCS Labs — Humanoid Control Systems株式会社. All rights reserved.