Token-World RoboTwin checkpoints

These are the reproduced RoboTwin checkpoints for Token-World: World Modeling in Vision-Language Model Token Space for Robot Manipulation. The code is available in the Token-World GitHub repository.

Artifacts

Component	File	Bytes
Three-view S-VAE	outputs/svae_d16_3views/best.pt	1,909,450,486
cam_high World Model, 100k updates	outputs/world_model_svae_d16_cam_high_100k/checkpoint_step_0100000.pt	1,498,034,990
Frozen-policy DeepStack adapter	outputs/deepstack_adapter_all50_20261031/adapter_best.pt	46,583,417
cam_high RGB decoder, epoch 500	outputs/rgb_decoder_svae_d16_cam_high/decoder_epoch_0500.pt	1,184,382,887

checkpoints.json records the SHA-256 checksum of every file. Files retain the original checkpoint payloads, including training metadata where present. Download them into the code checkout to preserve the relative paths expected by the launchers.

Training and evaluation contract

The S-VAE compresses each of 108 Qwen visual tokens from 2,560 to 16 channels, retaining a 9 × 12 spatial grid. It was trained using cam_high and both wrist cameras. The World Model uses cam_high only and predicts at 50 Hz. The reproduced recurrent protocol keeps attention KV inside each segment and passes only the GRU hidden state across segment boundaries.

The fixed RoboTwin split has 4,499 training and 500 validation episodes across 50 tasks. The balanced100 open-loop panel contains one success and one failure per task. It scores 22,152 predicted frames after real-frame-zero conditioning, using recorded actions to drive generation. The native action comparison uses the adapter plus the frozen QwenOFT policy on predictions, with direct real-RGB policy outputs as reference.

Metric	Pooled micro score
Merger feature cosine	0.773141
Merger feature NMSE	0.457262
Native-interface action cosine	0.959835
Native-interface action NMSE	0.078208

The native action protocol differs from the historical merger-only paper evaluation. See the code README for both results and their definitions.

Runtime and dependencies

The tested environment is Python 3.12 on PPU with SDK PyTorch 2.6.0. The repository's installer creates a checkout-relative venv using the SDK's system packages. Qwen3-VL-4B-Instruct, the frozen RoboTwin QwenOFT policy bundle, raw RoboTwin data or exported latent caches, and simulator assets are separate dependencies. These four files contain the Token-World components.

The RGB decoder can distort object shapes and colors. It is the epoch-500 checkpoint used in the reproduced visual evaluations. Experimental decoder replacements have separate diagnostic results.

Consult the code repository's license and third-party notices alongside the separate terms of downloaded backbones, policies and datasets.

Download and verify

Clone the code repository, install its environment, and run one of the following from the checkout.

# Hugging Face
bash scripts/run_python.sh -m pip install 'huggingface-hub>=0.30,<1'
bash scripts/run_python.sh scripts/download_checkpoints.py --repo-id ChuyaoFu/Token-World

# ModelScope
bash scripts/run_python.sh -m pip install 'modelscope-hub==0.4.3'
bash scripts/run_python.sh scripts/download_checkpoints.py \
  --provider modelscope --repo-id Zellovo/Token-World --revision master

Hugging Face mirror · ModelScope mirror

Optional QwenOFT policy bundle

The original RoboTwin cam_high step-50,000 policy is also provided. policy.json pins a 9,785,132,794-byte full policy checkpoint and a small support archive containing the exact StarVLA/RoboTwin source snapshots, primary configurations, action statistics, tokenizer metadata and licenses.

From the code checkout:

bash scripts/run_python.sh scripts/download_policy.py --provider modelscope
# Or --provider huggingface
export POLICY_BUNDLE="$PWD/outputs/policies/robotwin_qwenoft"
export QWEN_MODEL=/path/to/Qwen3-VL-4B-Instruct

The frozen base model is obtained upstream. The policy weight is Apache-2.0, its StarVLA source is MIT, and included RoboTwin files retain their original license. See LICENSES.md; the repository's MIT file does not relicense these components.

Data and derived caches

The exact data export already exists at Zellovo/robotwin-success-fail-lerobot: 4,999 trajectories and 14,997 videos. ModelScope currently requires authenticated access. Follow the code repository's docs/data.md for download instructions.

Latent caches and training statistics are generated locally with scripts/reproduce_world_model.sh --stage export; they are not missing uploads. Baseline checkpoints are outside the Token-World release scope.