lawam-actionvae-models

LaWAM / ActionVAE experiment checkpoints

This repository contains selected locally trained artifacts needed to resume the ActionVAE and LaWAM experiments:

• LIBERO and RoboTwin Effect/LaWM and ActionAE teachers;
• the local official-config LIBERO Stage-2 baseline reproduction;
• the selected release-residual policy;
• the selected Phase-A, Phase-B, and Phase-C policies from the 50k recipe.

It intentionally does not duplicate unmodified Qwen, DINO, or official LaWAM release checkpoints. Use migration/scripts/restore_public_models.sh from the companion source repository to restore those models at pinned revisions.

Licenses and provenance

This is a mixed-license research artifact collection. The policy lineage uses Qwen3-VL (Apache-2.0) and the MIT-licensed LaWAM release. Effect/LaWM artifacts use DINOv3 materials and remain subject to the upstream DINOv3 License included at licenses/DINOv3_LICENSE.md, including its redistribution, trade-control, and prohibited-use terms. Downstream users are responsible for complying with the licenses of all base models and datasets used to produce these checkpoints.

The repository contains model states and compact training metadata. It does not contain raw LIBERO or RoboTwin demonstrations.