yuanty/G07-preview-RoboDojo

G07-preview — RoboDojo evaluation submission

G07-preview is Galaxea AI’s autoregressive world-action model built on Cosmos3-Nano. This repository provides an eval-only XPolicyLab adapter, the evaluated step_020904 checkpoint and frozen inference configuration, and local evaluation evidence.

Install the adapter

Download the small code and submission files (pip install huggingface_hub==0.29.2):

from huggingface_hub import snapshot_download
snapshot_download(
    "yuanty/G07-preview-RoboDojo",
    allow_patterns=["G07_preview/*", "submission/*"],
    local_dir="g07-preview",
)

Place g07-preview/G07_preview in your RoboDojo workspace at XPolicyLab/policy/G07_preview/, then follow the policy README. Its download_weights.sh installs the pinned model assets into the layout consumed by the adapter.

Alternatively, download G07-preview-eval-only.tar.gz and verify it against the SHA-256 file. The archive contains the adapter and submission evidence; model weights are retrieved by download_weights.sh. File hashes and exact asset revisions are listed in MANIFEST.json.

Submission materials

• Method description.
• Eval-only submission description.
• Adapter validation.
• Complete local results, with raw per-layout results and provenance in the same directory.

The local results cover three seeds and 6,300 episodes. The official summarizer reports 41.46 ± 0.36 (population standard deviation), with 100% coverage. Their consistency with the submitted adapter is documented in the validation report. Official leaderboard scores are determined by the RoboDojo evaluation system.

Licenses

The adapter code in G07_preview/ and the G07-preview checkpoint in checkpoints/robodojo/ are released under the Apache License 2.0 (G07_preview/LICENSE). Five vendored modules are derived from transformers, diffusers and DiffSynth-Studio, as listed in G07_preview/THIRD_PARTY_NOTICES.md.

Model assets and third-party licenses

checkpoints/ holds every file the adapter loads, in the layout download_weights.sh installs:

Path	Source	License
checkpoints/robodojo/	G07-preview checkpoint step_020904 and frozen inference configuration, fine-tuned by Galaxea AI from Cosmos3-Nano	Apache-2.0 (LICENSE, NOTICE); keeps the OpenMDW-1.1 copy required by Cosmos3-Nano (LICENSE-Cosmos3-Nano)
checkpoints/Cosmos3-Nano/	Unmodified transformer/, text_tokenizer/, vision_encoder/ from nvidia/Cosmos3-Nano at 411f42a8fdfb8c5b2583cb8786e0938f49796eaa, with NVIDIA's model card	OpenMDW-1.1 (LICENSE)
checkpoints/DiffSynth-Studio/Wan-Series-Converted-Safetensors/	Unmodified Wan2.2_VAE.safetensors from DiffSynth-Studio, converted from the Wan2.2 VAE	Apache-2.0 (LICENSE, NOTICE)

Release commitment

Within three days after this entry is listed on the RoboDojo leaderboard, we will publicly release the training code, evaluated checkpoint, and a technical blog. The release through XPolicyLab will also include deployment code, configuration files, and reproduction instructions.