Foundation Action reproduction checkpoints
This repository contains the exact checkpoints and loader-adjacent metadata used by the Kesai Labs Foundation Action reproduction.
The bundle covers:
- DINO temporal dynamics step 100,000, its LingBot teacher, action/state and trajectory normalizations, depth/MoGe auxiliaries, and evaluation probe;
- corrected-DA3 dynamics step 6,000, DA3-Giant and GAM base weights, experiment metadata, calibration artifacts, and both evaluation probes.
Every file is bound by size, SHA-256, and role in
checkpoint-manifest.json. Verify the manifest before loading any pickle-based
.pt or .pth file. These files are intended for the locked-uv workflow in
kesai-labs/foundation-action
at validated code revision eac516fa40519a50a1249388cc61484db8bf1742.
That exact code is also included as
source/foundation-action-eac516f.tar.gz with SHA-256
e4ac407d7b7826bd42f2afeffa31b4920a24222638798462c91f0c0948ec504c.
The full 1,024-parent Alpamayo1.5 K=6 and historical Top-10 evidence is recorded
in docs/2026-07-22-ar15-full-reproduction.md in that code repository.
The README section “Download-to-retrieval example” is the smaller supported
PAI-derived GT+9 path and explicitly does not claim historical metric identity.
No PAI data, camera/egomotion archive, frozen candidate, evaluator source, or golden result is included.
License and access
The bundle combines artifacts with different upstream terms:
- LingBot-VLA2/DINO-Video and LingBot-Depth: Apache License 2.0;
- MoGe-2 and Geometric Action Model: MIT;
- DA3 source code: Apache License 2.0;
- DA3-GIANT weights and derivatives: CC BY-NC 4.0, as identified by the upstream DA3 model table.
Corresponding Apache/MIT license texts are under licenses/. The bundled
Foundation Action source is MIT-licensed and contains its own notices.
DA3-GIANT and
the corrected-DA3 derivative checkpoint are non-commercial artifacts; users
must preserve attribution and comply with CC BY-NC 4.0. This repository
contains no PAI data. The license: other marker reflects the mixed bundle and
does not replace any upstream license or access condition.