---
title: Abot-PhysWorld
canonical_url: "https://www.modelscope.cn/models/amap_cvlab/Abot-PhysWorld"
md_url: "https://www.modelscope.cn/models/amap_cvlab/Abot-PhysWorld.md"
repository: amap_cvlab/Abot-PhysWorld
last_updated: 2026-05-21
license: "Apache License 2.0"
pipeline_tag: text-to-video-synthesis
tasks:
  - text-to-video-synthesis
parameters: 32.8B
tensor_type:
  - BF16
library_name:
  - pytorch
  - safetensors
frameworks:
  - Pytorch
language:
  - en
  - zh
downloads: 1072
stars: 3
tags:
  - text-to-video
  - diffusion
  - robotics
  - world-model
  - embodied-ai
  - sft
---

# Abot-PhysWorld

> Abot-PhysWorld - amap_cvlab 在 ModelScope 开源的模型。Checkpoint for https://github.com/amap-cvlab/ABot-PhysWorld

amap_cvlab/Abot-PhysWorld 是 ModelScope 魔搭社区上的 32.8B 参数text-to-video-synthesis模型，采用 Apache License 2.0 许可。

- **Repository**: amap_cvlab/Abot-PhysWorld
- **License**: Apache License 2.0
- **Tasks**: text-to-video-synthesis
- **Parameters**: 32.8B
- **Tags**: text-to-video, diffusion, robotics, world-model, embodied-ai, sft
- **Downloads**: 1072
- **Stars**: 3
- **Last updated**: 2026-05-21

Source: https://www.modelscope.cn/models/amap_cvlab/Abot-PhysWorld

---

# ABot-PhysWorld 

This repository releases the checkpoint from **ABot-PhysWorld**, a physically consistent and action-controllable video world model for robotic manipulation.

## Model Details

- **Model name**: ABot-PhysWorld
- **Organization**: AMAP CV Lab
- **Architecture**: 14B Diffusion Transformer (DiT) world model
- **Task**: Text/action-conditioned robotic manipulation video generation
- **Released checkpoint**: TI2V checkpoint
- **Training stage for this checkpoint**: SFT fine-tuning

## Intended Use

Use this model for:

- Physics-aware robotic manipulation video generation
- Zero-shot scenario testing for embodied world models
- Action-controllable video generation research
- Benchmark evaluation such as PBench/EZS-Bench style protocols

## Out-of-Scope Use

- Safety-critical real robot control without downstream verification
- Human identity-sensitive or misleading synthetic media generation
- Any usage that violates law, platform policy, or license terms

## Training Notes

- **Data pipeline**: Large-scale manipulation data curation with motion/semantic/action consistency filtering
- **Optimization recipe**: Physics-aware preference optimization and action-conditioning modules (project-level training framework)
- **Primary domain**: Robotic manipulation with physical consistency emphasis

For full pipeline details, please refer to the project repository:

- [ABot-PhysWorld GitHub](https://github.com/amap-cvlab/ABot-PhysWorld)

## Quick Start (Example)

Minimal loading example:

```python
import torch
from safetensors.torch import load_file

ckpt_path = "xxx.safetensors"
state_dict = load_file(ckpt_path)

# TODO: build your ABot-PhysWorld model architecture, then load state_dict
# model = YourVideoModel(...)
# model.load_state_dict(state_dict, strict=False)
# model.eval().to("cuda")
```

If you are using the DiffSynth-Studio workflow, set model path to:

`path/xxx.safetensors`

## Evaluation Snapshot

Reported project-level results:

| Capability | Benchmark | Ours | Best Baseline | Gain |
|---|---|---|---|---|
| Physical Fidelity | PBench | **0.8491** | 0.8096 (Wan2.5) | +3.95% |
| Zero-Shot Generalization | EZS-Bench | **0.8366** | 0.7951 (WoW) | +4.15% |
| Action Control | Trajectory Consistency | **0.8522** | 0.8157 (Enerverse) | +3.65% |

## Limitations

- Performance can degrade under extreme domain shift or very long-horizon tasks
- Generated videos may still contain occasional temporal or contact artifacts
- This checkpoint is for research usage and should not be treated as guaranteed physical simulation

## Ethical Considerations

- Generated videos can be plausible but not physically guaranteed in all cases
- Human review is required before using outputs for decisions or claims
- Do not use outputs to overstate real-world robotic capability or safety

## Citation

If you use ABot-PhysWorld or this checkpoint, please cite:

```bibtex
@article{
  title={ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment},
  author={Yuzhi Chen, Ronghan Chen, Dongjie Huo, Yandan Yang, Dekang Qi, Haoyun Liu, Tong Lin, Shuang Zeng, Junjin Xiao, Xinyuan Chang, Feng Xiong, Xing Wei, Zhiheng Ma, Mu Xu},
  year={2026}
}
```

## License

This model is licensed under the **Apache 2.0 License**.

You must not use this model in a manner that infringes, misappropriates, or violates any third party’s rights, including intellectual property rights.

## Acknowledgement

This project builds upon:

- [Wan2.1](https://github.com/Wan-Video/Wan2.1)
- [VACE](https://github.com/ali-vilab/VACE)
- [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio)

## 中文说明

该仓库发布 **ABot-PhysWorld** 的一个 SFT 阶段权重，用于机器人操作视频世界模型研究。  
模型强调物理一致性、动作可控性和零样本泛化能力，适用于学术研究与基准评测。