AR-GO checkpoints
This repository contains the 2026-07-29 reproducibility snapshot for AR-GO.
Verified packages
| Path | Initialization | LIBERO evidence |
|---|---|---|
spatial-ar-go-93/ | generic VLM → clean-start Pro 10k → bounded AR-GO | Spatial 93/100 |
spatial-cleanstart-base-10k/ | generic VLM | exact AR-GO training base |
future-prior-stage1/ | Qwen3-VL clean-start Stage 1 | frozen future prior |
object-cleanstart-20k-100/ | generic VLM | Object 100/100 |
spatial-cleanstart-150k-91/ | generic VLM | Spatial 91/100 |
No task-finetuned public VLA-Adapter checkpoint initializes the released AR-GO model. See the GitHub run contract and raw JSON/JSONL evidence for the complete provenance and claim boundary.
Download
hf download Jaqen0-0/AR-GO \
--local-dir ./AR-GO-checkpoints
Integrity
Per-file SHA-256 manifests are mirrored in the GitHub repository under
weights/manifests/. The future-prior Stage-1 file must hash to:
8238e44578d5d9c6d609cd82e6f220c315e065d498b4407a4ab6e8d27ce88b92
Important
The 93/100 value is a completed strict LIBERO-Spatial evaluation. It is not a claim that all paper tables are reproduced. The clean-start object and spatial checkpoints are released as strong controls and reproducible policy bases.
Foundation models, datasets, and simulator assets are not redistributed here and retain their upstream licenses.