InternW0-RoboTwin-clean

Model description

InternW0-RoboTwin-clean is the task-posttrained InternW0 checkpoint for the RoboTwin 2.0 Clean2Random setting. It is initialized from the heterogeneous InternW0-Base pretraining checkpoint and posttrained only on clean demonstrations. The selected checkpoint is then evaluated on clean scenes and unseen randomized scenes.

InternW0 couples a slow Wan2.2-TI2V-5B VideoDiT planner with a fast ActionDiT executor. Observation-conditioned chunk K/V routing adapts cached future-video context to each newly observed state, allowing the model to reuse long-horizon predictions while generating responsive action chunks.

This is a native w0 checkpoint, not a standalone Transformers model. It is intended for RoboTwin evaluation or compatible continued posttraining.

Files and integrity

File	Purpose	Size	SHA-256
InternW0-RoboTwin-clean.pt	Canonical model-only native checkpoint	14,705,453,767 bytes	b718314f57c2febb159deb03d92cf9ebba830f7c75920c0b6e34c4903b91f5a2
pytorch_model.bin	Zero-copy LFS alias used for PyTorch download statistics	Same LFS object	b718314f57c2febb159deb03d92cf9ebba830f7c75920c0b6e34c4903b91f5a2
config.json	Machine-readable model/download-counting manifest	1,702 bytes	82d8a7a6f25fe14720dc7d6e02136fa02fff0f0b2650a29866eacc4edf275347
checkpoint_metadata.json	Embedded behavior/training/statistics/provenance export	16,176 bytes	6e4b67d29ca9a369c870f510e9304ffb3e4aec974f6e38a09c35d598c4fbaf54
normalization.json	Matching RoboTwin action/proprio statistics	4,014 bytes	31362a24460803c5080ef925b807acbc7af3e08d847c9db222bc2a33c0d96653
posttrain_config.yaml	Portable public-code posttraining config	1,661 bytes	109ca7e501c65e4072060ec7477da2a5486f45aed095ece661714ef92f477f88
evaluation_config.yaml	Portable public-code evaluation config	885 bytes	89545dbed7a002c3a1e09edf48a3540d30d977e96b89e6fad64ab0b92020fd38
source_training_config.yaml	Historical resolved config with portable paths	6,175 bytes	6455e1dda379528fe76d5c85e4161bce651286897567bd55ebe2c45c7e9e4312
source_evaluation_config.yaml	Historical evaluation reference with portable paths	7,312 bytes	6b16dc8583a79d69a253f1b3a367aab072c8dcf24737ece784c9a926d807aff3
paper_results.json	Paper-reported Clean2Random result record	621 bytes	9396065aded7413f6d4ce0228ac512bb3c97fa0b32157fb8b15f19c1e82ef2cf
PROVENANCE.md	Selection, lineage, training, and evaluation provenance	2,446 bytes	0cca6d854f5bf6ee3dccd84128584bdae398f9f02e053c016d9d097bd944680a
LICENSE	MIT license	1,205 bytes	b1d3d0160bdd9748bfc4e8acb9ab8fbce7ab725c827fd6fc673261f57aae0306

pytorch_model.bin is a second Git-LFS path to the exact same object as InternW0-RoboTwin-clean.pt. It does not duplicate the 14.7 GB payload.

Checkpoint summary

• Selected step: 10545
• Serialization dtype: bfloat16
• Model-only checkpoint: yes
• Optimizer state: excluded
• Tensors: 1,977
• Parameter elements: 7,352,528,643
• Source heterogeneous pretraining step: 30000
• Target domain: robotwin_wam/aloha
• Target soft prompts: K=8
• Canonical action/proprio dimensionality: 37
• Valid ALOHA joint/gripper slots: 14
• Action horizon: 64
• Action chunk size: 16
• Video plan horizon: 96 action ticks
• Video/action stride: 8
• History frames: 0
• Action semantics: native absolute joint command at interval end

The checkpoint embeds a schema-v4 behavior_config, training_recipe, posttrain_statistics, and posttrain_provenance. The public loader performs a semantic contract check before loading weights.

RoboTwin Clean2Random training

The retained recipe uses:

• clean demonstrations only;
• native RoboTwin joint-space data;
• 25 Hz sampling (sampling_stride=2 from 50 Hz);
• every valid native window;
• ColorJitter with brightness 0.2, contrast 0.2, saturation 0.2, and hue 0;
• action offset 0;
• batch size 12 per process;
• 15 epochs;
• learning rate 3e-5 with cosine decay;
• seed 42.

The matching normalization statistics are in normalization.json. The historical run resumed from step 5,000 of the preceding clean run and selected step 10,545 from the continued best Clean2Random run. See PROVENANCE.md for the full lineage.

Evaluation result

The paper-reported RoboTwin 2.0 Clean2Random results are:

Split	Success rate
Clean	83.20%
Randomized, unseen during training	68.00%
Average	75.60%

The source evaluation reference uses 50 tasks, 15 episodes per task, 10 inference steps, a 64-action horizon, four action chunks per video prefill, and a fixed-period 25 Hz controller. The historical resolved protocol is preserved in source_evaluation_config.yaml.

Required external assets

This checkpoint is not self-contained. Users must separately obtain and comply with the licenses for:

hf download Wan-AI/Wan2.2-TI2V-5B --local-dir assets/wan22
hf download facebook/dinov3-vitb16-pretrain-lvd1689m \
  --local-dir assets/dinov3-vitb16-pretrain-lvd1689m

A RoboTwin checkout, the matching RoboTwin data export, and its UMT5 text cache are also required. The repository does not bundle datasets, simulation code, Wan2.2 files, or DINOv3 weights.

Quick start

1. Install the public codebase

git clone https://github.com/InternRobotics/InternW0.git
cd InternW0

conda create -n w0 python=3.10 -y
conda activate w0
pip install -U pip setuptools wheel
pip install torch==2.7.1 torchvision==0.22.1 \
  --extra-index-url https://download.pytorch.org/whl/cu128
pip install -e .

2. Download this checkpoint

hf download InternRobotics/InternW0-RoboTwin-clean \
  config.json pytorch_model.bin checkpoint_metadata.json \
  normalization.json posttrain_config.yaml evaluation_config.yaml \
  source_training_config.yaml source_evaluation_config.yaml \
  paper_results.json PROVENANCE.md LICENSE \
  --local-dir assets/robotwin

mv assets/robotwin/pytorch_model.bin \
   assets/robotwin/InternW0-RoboTwin-clean.pt

3. Run RoboTwin evaluation

export W0_WAN_MODEL_DIR="$PWD/assets/wan22"
export W0_DINO_WEIGHTS="$PWD/assets/dinov3-vitb16-pretrain-lvd1689m"
export W0_ROBOTWIN_DATA="$PWD/data/robotwin"
export W0_ROBOTWIN_STATS="$PWD/assets/robotwin/normalization.json"
export W0_ROBOTWIN_CHECKPOINT="$PWD/assets/robotwin/InternW0-RoboTwin-clean.pt"
export ROBOTWIN_ROOT="$PWD/external/RoboTwin"

python experiments/robotwin/eval_robotwin_single.py \
  EVALUATION.robotwin_root="$ROBOTWIN_ROOT" \
  EVALUATION.task_name=<task> \
  EVALUATION.ckpt="$W0_ROBOTWIN_CHECKPOINT"

For compatible continued posttraining:

bash scripts/train_zero1.sh 8 --config-name posttrain_robotwin

Download statistics

Hugging Face uses library-specific query files for model download statistics. For the pytorch library, the Hub counts HTTP GET and HEAD requests to pytorch_model.bin. This repository provides that path as a zero-copy alias of the canonical checkpoint. The badge above reads the public Hub API downloads field, corresponding to downloads in the last 30 days. Hub analytics are aggregated and may lag briefly.

Intended uses

• RoboTwin 2.0 Clean2Random reproduction and extension.
• Research on world-action models, context routing, and cross-embodiment robot learning.
• Compatible continued posttraining with the public w0 codebase.

Limitations

• This is a simulation-domain research checkpoint, not a deployed robot controller.
• It requires the public InternW0 codebase, external Wan2.2 and DINOv3 assets, RoboTwin data, text caches, and substantial compute.
• It is not directly compatible with transformers.AutoModel.
• Real-robot use requires independent safety analysis, constrained workspaces, and human oversight.

License

The weights and metadata in this repository are released under the MIT License. Third-party assets obtained separately remain subject to their original licenses.

Citation

If you use InternW0, please cite:

@article{internw0,
  title = {InternW0: A Foundational Physical World Model for Efficient Real-World Interactions},
  author = {Jisong Cai and Yao Mu and Ganlin Yang and Zhe Cao and Zhangzheng Tu and Xing Gao and Kailin Li and Xinyu Zhan and Lixin Yang and Yangkun Zhu and Haoxiang Ma and Ming Zhou and Qiaojun Yu and Yufei Xue and Liqun He and Yifei Yao and Yifan Zhu and Long Ling and Bingqi Jiang and Haoyu Guo and Xueyue Zhu and Bowen Zhou and Bin Zhao and Tianfan Xue and Chunhua Shen and Weinan Zhang},
  journal = {arXiv preprint arXiv:2609.27656},
  year = {2026}
}