RSCI checkpoints

Full 40,000-step training checkpoints for Rethinking State Conditioning in Vision-Language-Action Models.

Layout

newcheckpoint/<architecture>/<model>/<task>/<vlm-mode>/<run>/

• Architecture: Psi0.
• Models: STATE, NO-STATE, VGQ-AE, (VGQ-AE-D); further variants retain their own names.
• Tasks: betweentables, hug, oven, faucet, xmovepick, handover.
• VLM mode: frozen or lora.

Each published run includes checkpoints/ckpt_40000/ with model weights, optimizer, scheduler and all distributed-rank random states, plus training configuration, launch contract and a SHA256 manifest. Runs and tasks are kept separate. Earlier intermediate checkpoints, training datasets and evaluation videos are not included.

The accompanying configurations identify the actual model, data and source version. Checkpoint publication does not assert an evaluation success rate. See each run's manifest for provenance and exact file hashes.