Skip to content

LICENSING.md

Latest commit

 

History

History
114 lines (80 loc) · 4.89 KB

LICENSING.md

File metadata and controls

114 lines (80 loc) · 4.89 KB
Raw

LICENSING — what may be published, and on what terms

Task M6-T2. WORKPLAN marks it blocking: "Publishing finetuned derivatives of NVIDIA-licensed checkpoints is redistribution of derived models. Confirm the terms, produce the attribution text, and document the base checkpoint and job config for every published weight set."

Decision taken by the project owner (Carlo Dentesano) on 2026-08-08: weights may be published.

This page records the terms that make that decision correct, so nobody has to take it on trust.

The terms

The upstream repository is dual-licensed, and its own LICENSE file at the repository root says so explicitly:

Component Licence
Source code, scripts, software Apache License 2.0
Trained model checkpoints and weights NVIDIA Open Model License (last modified 2025-10-24)

The NVIDIA Open Model License is permissive about exactly the thing M6-T2 was worried about. Section 2.2 grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, revocable license to … create derivative works of … distribute and import the Model", and Section 2.4 states plainly: "You own Your Derivative Models."

So a finetuned GEAR-SONIC checkpoint is ours to publish. The blocker was real — it just resolves in our favour once read.

What we must do when we publish weights

Section 3 imposes three conditions, and only three:

  • (a) Ship a copy of the Agreement. Include the upstream LICENSE file alongside any published weight set.

  • (b) Carry the attribution notice, verbatim:

    Licensed by NVIDIA Corporation under the NVIDIA Open Model License.

    It may sit wherever other third-party notices live.

  • (c) Cosmos models only — a phrase "Built on NVIDIA Cosmos". See the open question below.

Section 2.1 adds two ways to lose the licence, both of which we should simply not do: initiating patent or copyright litigation over the Model, and circumventing built-in safety guardrails without comparable alternatives. Neither is in scope for a boxing game.

Section 2.3 requires use in accordance with NVIDIA's Trustworthy AI terms.

⚠ One open question, and it is small

Condition (c) applies to a "NVIDIA Cosmos Model", defined in §1.4 as "a multimodal Model that is covered by this Agreement". GEAR-SONIC is a whole-body control policy, not a multimodal model, so on a plain reading (c) does not apply and the phrase "Built on NVIDIA Cosmos" would be inaccurate rather than merely redundant.

That is my reading, not a lawyer's. It costs nothing to check with NVIDIA before the first public weight release, and it is the only point on this page where the answer is not simply written down.

Attribution text to ship

Put this in the release notes, the model card, and any page offering a download:

This model is a derivative of GEAR-SONIC from NVlabs/GR00T-WholeBodyControl.
Licensed by NVIDIA Corporation under the NVIDIA Open Model License.

Base checkpoint : gear_sonic_deploy/policy/release
                  model_encoder.onnx  sha256 013ab0287236aa27...
                  model_decoder.onnx  sha256 c7241a123eaa36b5...
Robot model     : g1_29dof_old.xml    sha256 58660a6f1d0d33ff...

OpenRoboxing source code is Apache-2.0, as is the upstream source it builds on.

The exact hashes for any given release come from that release's season manifest (league/manifest.py); the ones above are illustrative and will differ per freeze.

Per-weight-set documentation

M6-T2 also requires "the base checkpoint and job config for every published weight set". The machinery exists and is not optional:

  • tools/freeze_season.py pins every asset by SHA-256, including both ONNX files and the robot model. A manifest is the release record.
  • A finetuned weight set additionally needs its job config, which is S-T2's to produce. Until S-T2 exists there are no finetuned weights to publish — only upstream's, unmodified, which this page already covers.

The gate in code

league/manifest.py refuses to mark a manifest released unless the caller passes an exact acknowledgement string. That gate stays, deliberately: this page makes the answer knowable, and the gate makes publishing a deliberate act rather than a default. Release a manifest with:

python -m openroboxing.tools.freeze_season \
    --season season-0 --at <iso8601> \
    --released M6-T2-signed-off \
    --out seasons/season-0.json

Status
Terms confirmed ✅ NVIDIA Open Model License permits derivative models and their distribution
Attribution text produced ✅ above
Human sign-off ✅ Carlo Dentesano, 2026-08-08
Cosmos clause checked with NVIDIA ⬜ open, low risk (see above)
Finetuned weight sets documented n/a — none exist yet (S-T2)

Nothing has been published. This page and the release gate say what may be; pushing files anywhere is a separate, deliberate act by a human.