InternRobotics / InternW0-Delta

A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data 
 English · 中文 
 🚀 Quick start ·
 🔥 Training ·
 🦾 Evaluation & deployment ·
 📦 Models 
 🌐 Project page ·
 📄 Paper / arXiv ·
 🤗 Hugging Face (coming soon)
## ✨ Overview
**InternW0-delta (InternW0-Δ)** is a unified world–action model that learns from robot and human demonstrations. It combines pretrained video dynamics, a frozen vision-language model, and training-only 4D supervision to generate robot actions.
- 🧩 **World–Action MoT.** A pretrained video expert and an action expert exchange information through masked self-attention, with visual memory and task-conditioned scene semantics.
- 🔮 **Causal Imprint.** Compact queries learn action-relevant scene changes from future supervision, making predictive features available to the action expert.
- 🌌 **4D distillation.** A frozen Track4World teacher transfers geometry and motion knowledge during training. The teacher and distillation branch are absent from action inference.
The repository covers data preparation, pretraining, 4D distillation, post-training, evaluation, and real robot deployment.
## 🚀 Quick start
### 1. Install
Use Linux, Python 3.11, a CUDA GPU, and FFmpeg. DeepSpeed extensions require a C++ compiler and a matching CUDA toolkit.
 [code] 
### 2. Prepare models and data
Download the backbones and policy weights following the model guide. For post-training, place the initialization weights at `checkpoints/pretrain.pt`.
Prepare the selected dataset using the data guide. Run the commands below from the repository root.
### 3. Post-train a policy
For LIBERO:
 [code] 
Continue to LIBERO Plus evaluation, or choose another task below.
## 🔥 Training
### Pretraining
Choose datasets through `source_files` in dataset.yaml. Training loads each selected source's normalization statistics automatically.
Prepare the action expert initialization, then run:
 [code] 
To continue from pretrained weights, add `resume=checkpoints/pretrain.pt model.skip_dit_load_from_pretrain=true`; action expert initialization is then unnecessary. For custom data, see normalization.
### 4D distillation
Select data, cache Track4World teacher features, and continue training with the 4D distillation guide.
### Task configurations
Use the task name with `python tools/text_cache.py task= ` and `bash run.sh `. See post-training data for the required exports.
| Task | Configuration | Evaluation / deployment |
| --- | --- | --- |
| LIBERO | `libero` | LIBERO Plus |
| RoboTwin | `robotwin` | RoboTwin |
| RoboDojo | `robodojo` | RoboDojo |
| Ebench | `ebench`, `ebench_joint_ee` | — |
| Real robots / RTC | `rtc` | Deployment |
## 🦾 Evaluation and deployment
Download the task weights, then follow the corresponding environment setup and run instructions:
| Guide | Use |
| --- | --- |
| LIBERO Plus | Evaluate LIBERO policies under task perturbations |
| RoboTwin | Post-train on clean demonstrations and evaluate clean → random |
| RoboDojo | Run th