Skip to content

Commit 6e311ce

Browse files
committed
Release ABot-World-0
1 parent c327d4d commit 6e311ce

99 files changed

Lines changed: 19762 additions & 191 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

.gitignore

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -211,3 +211,4 @@ __marimo__/
211211
assets/video-ori
212212
assets/video
213213
compress.sh
214+
website.sh

NOTICE

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,7 @@
1+
ABot-World
2+
Copyright 2026 AMAP CV LAB
3+
4+
This product includes software developed by AMAP CV LAB.
5+
6+
This product includes or adapts third-party software components. See
7+
THIRD_PARTY_NOTICES.md for attribution and license references.

README.md

Lines changed: 83 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -1,49 +1,105 @@
11
# 🌍 ABot-World: Real-Time Interactive World Simulation on a Single Desktop GPU
22

33
[![Project](https://img.shields.io/badge/🌐%20%20Project-ABot%20%20World-blue.svg)](https://amap-cvlab.github.io/ABot-World/)
4+
[![Studio](https://img.shields.io/badge/🎮%20%20Studio-ABot%20%20World%20%20Studio-green.svg)](https://abot-world.amap.com)
45
[![Paper](https://img.shields.io/badge/Arxiv-Coming_Soon-red)](#)
5-
[![Code](https://img.shields.io/badge/Code-Coming_Soon-181717.svg?logo=GitHub)](https://github.com/amap-cvlab/ABot-World)
6-
[![Model](https://img.shields.io/badge/Weights-Coming_Soon-yellow)](#)
6+
[![Code](https://img.shields.io/badge/Code-GitHub-181717.svg?logo=GitHub)](https://github.com/amap-cvlab/ABot-World)
7+
[![Model](https://img.shields.io/static/v1?label=%F0%9F%A4%97%20Model&message=HuggingFace&color=yellow)](https://huggingface.co/acvlab/ABot-World-0-5B-LF)
8+
[![Model](https://img.shields.io/static/v1?label=%F0%9F%A4%96%20Model&message=ModelScope&color=purple)](https://modelscope.cn/models/amap_cvlab/ABot-World-0-5B-LF)
79

8-
> **TL;DR:** ABot-World is a next-generation Action-Conditioned World Model that achieves **real-time interactive inference on a single consumer-grade Desktop GPU**.
9-
10-
We are currently preparing the technical report, code, and model weights for public release. In the meantime, explore our interactive demos below.
10+
> **TL;DR:** ABot-World turns a single NVIDIA RTX 5090 desktop GPU into a real-time interactive world simulator, enabling infinite action-conditioned world rollout at 720P, 16 FPS, 1.2s latency, and 19GB GPU memory.
1111
1212
## 🚀 Key Highlights
1313

14-
***Real-Time Desktop Inference:** Breaking the latency barrier. Through highly optimized inference architecture and compute-efficient scaling, ABot-World enables fluid, real-time rollout generation on a single desktop GPU. You no longer need a data center to simulate the world.
15-
* 🧠 **Native Spatial Intelligence**: To capture the infinite variations of the physical world, ABot-World extracts robust spatial structures and complex dynamics implicitly. By natively embedding physical priors into its generation process, it seamlessly unifies visual rendering and physical simulation within the latent space.
16-
* ♾️ **Long-Horizon Consistency:** Exposure bias and compounding errors have long plagued auto-regressive generation. By introducing a novel Parametric Memory module that acts as a global state sink, ABot-World maintains strict spatial topology and physical consistency over extraordinarily long rollouts.
17-
* 🏆 **Top-Tier Physical Reasoning:** Evaluated against state-of-the-art models, ABot-World achieves top-tier performance on industry-standard benchmarks including [**VBench 2.0**](https://huggingface.co/spaces/Vchitect/VBench_Leaderboard) (annouced by [ABot-World v0.1](https://github.com/amap-cvlab/ABot-World)), [**WorldArena**](https://huggingface.co/spaces/WorldArena/WorldArena) & [**GigaBran**](https://huggingface.co/spaces/open-gigaai/CVPR-2026-WorldModel-Track-LeaderBoard) (announced by [ABot-PhysWorld](https://huggingface.co/spaces/WorldArena/WorldArena)) and [**WorldScore**](https://huggingface.co/spaces/Howieeeee/WorldScore_Leaderboard) (announced by [FantasyWorld 1.0](https://fantasy-amap.github.io/fantasy-world/)), particularly excelling in physics adherence and dynamic motion quality.
14+
* 🎮 **Action-Driven World Control:** Responds to user actions in real time, enabling continuous exploration instead of passive video playback.
15+
***Real-Time Desktop Inference:** Runs at 720p and 16 FPS on a single NVIDIA RTX 5090 desktop GPU, with 1.2s latency and 19GB GPU memory.
16+
* ♾️ **Infinite World Rollout:** Supports open-ended interactive world generation beyond fixed video-length limits.
17+
* 🧠 **Open-Ended World Imagination:** Expands the world with new scenes and dynamics during rollout, avoiding scene lock-in, without prompt switching.
18+
19+
## 📢 News
20+
- 2026-07-09: We release the causal student model `ABot-World-0-5B-LF`, inference code, our local gradio demo and online playground [ABot World Studio](https://abot-world.amap.com).
21+
22+
## 🛠️ Setup
23+
24+
> This installation was tested on: Ubuntu 22.04, CUDA 13.3, NVIDIA RTX 5090.
25+
26+
1. Clone the repository:
27+
28+
```bash
29+
git clone https://github.com/amap-cvlab/ABot-World.git
30+
cd ABot-World
31+
```
32+
33+
2. Install dependencies using conda:
34+
35+
```bash
36+
conda create -n aworld python=3.12 -y
37+
conda activate aworld
38+
pip install -r requirements.txt
39+
```
40+
41+
3. Download checkpoints:
42+
43+
```bash
44+
pip install -U "huggingface_hub"
45+
hf download <HF_REPO_ID> --local-dir checkpoints
46+
```
47+
48+
After downloading, the project should have the following checkpoint structure:
49+
50+
```text
51+
checkpoints/
52+
└── ABot-World-0-5B-LF/
53+
├── Wan2.2_VAE.pth
54+
├── taew2_2.pth
55+
├── models_t5_umt5-xxl-enc-bf16.pth
56+
├── diffusion_pytorch_model.safetensors
57+
└── google/umt5-xxl/
58+
```
59+
60+
The checkpoint paths are configured in `configs/long_forcing_dmd.yaml` and
61+
`configs/default_config.yaml`. The distilled generator weights are already
62+
merged into `ABot-World-0-5B-LF/diffusion_pytorch_model.safetensors`.
1863

19-
## 🕹️ Interactive Demonstrations
64+
## 🤗 Gradio Demo
2065

21-
### 1. Open World Exploration
22-
Experience fluid and consistent generation as you interact with the environment. Whether navigating via first-person camera adjustments or driving a character through diverse scenes, the simulation responds dynamically to your inputs. The model natively understands spatial depth, object permanence, and complex environmental interactions, delivering an immersive experience from any viewpoint.
66+
1. Run the demo online: 此处预留hf demo 链接
67+
2. Run it locally:
2368

24-
*<img src="assets/gifs/open-world-exploration-main.gif" width="800" alt="More Open World Exploration Demo">*
69+
```bash
70+
bash web_client/run.sh
71+
```
2572

26-
#### More Demos
27-
*<img src="assets/gifs/open-world-exploration.gif" width="800" alt="More Open World Exploration Demo">*
73+
Select a GPU with:
2874

29-
### 2. Long-Horizon World Exploration via a single Desktop GPU
30-
A continuous, uninterrupted rollout showcasing the power of our Parametric Memory. Through a highly compute-efficient architecture, ABot-World sustains strict global state topology and physical consistency over extended durations, generating expansive environments locally on a single Desktop GPU.
75+
```bash
76+
CUDA_ID=0 bash web_client/run.sh
77+
```
3178

32-
*<img src="assets/gifs/long-horizon.gif" width="800" alt="Long Horizon World Exploration Demo">*
79+
## License
3380

34-
🔗 **[Explore our previous work FantasyWorld](https://fantasy-amap.github.io/fantasy-world/)**
81+
This project is released under the Apache License 2.0. See `LICENSE`, `NOTICE`,
82+
and `THIRD_PARTY_NOTICES.md` for copyright and third-party attribution details.
3583

36-
### 3. Embodied Interaction & Physics Alignment (ABot-PhysWorld)
37-
A specialized variant of our foundation model, **ABot-PhysWorld**, focuses on robotic manipulation. It uses a novel DPO-based post-training framework with decoupled discriminators to actively suppress unphysical behaviors like object penetration and anti-gravity motion. A parallel context block enables precise spatial action injection for cross-embodiment control.
84+
## 🤝 Acknowledgement
3885

39-
🔗 **[Explore the ABot-PhysWorld Repository](https://github.com/amap-cvlab/ABot-PhysWorld)**
86+
This project builds on and is inspired by the following open-source projects: [Causal Forcing](https://github.com/thu-ml/Causal-Forcing), [AngelSlim](https://github.com/tencent/AngelSlim), [LightX2V](https://github.com/ModelTC/LightX2V), [taehv](https://github.com/madebyollin/taehv), [Wan2.2](https://github.com/Wan-Video/Wan2.2), [Helios](https://github.com/PKU-YuanGroup/Helios), from which the optimized Triton RoPE and normalization kernels in `wan/modules/helios_kernels` are derived.
4087

4188
## 🗓️ Roadmap
42-
- [x] Teaser & Demo Release
89+
- [x] Interactive Web Playground (ABot World Studio)
90+
- [x] Inference Code Release
91+
- [x] Local Gradio Demo Release
92+
- [x] Causal Student Model Release
93+
- [ ] Bidirectional Teacher Model Release
4394
- [ ] Technical Report (Arxiv)
44-
- [ ] Inference Code Release
45-
- [ ] Model Weights (Open Source)
46-
- [ ] Interactive Web Playground
4795

4896
## 📝 Citation
49-
If you find our work helpful, please stay tuned for our upcoming paper.
97+
If you find our work helpful, please cite our paper:
98+
99+
```
100+
@article{abot-world-0,
101+
title={ABot-World-0: Real-Time Interactive World Simulation on a Single Desktop GPU},
102+
author={ABot-World Team},
103+
year={2026}
104+
}
105+
```

THIRD_PARTY_NOTICES.md

Lines changed: 47 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,47 @@
1+
# Third-Party Notices
2+
3+
Unless otherwise noted, ABot-World source code is distributed under the Apache
4+
License, Version 2.0. This file summarizes third-party components that are
5+
included, adapted, or required by this repository.
6+
7+
## Wan / Wan2.2
8+
9+
- Upstream: https://github.com/Wan-Video/Wan2.2
10+
- License: Apache License 2.0
11+
- Copyright notice retained in source files:
12+
`Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.`
13+
- Local usage: model, attention, text encoder, tokenizer, VAE, and related Wan
14+
modules under `wan/`.
15+
16+
## AngelSlim
17+
18+
- Upstream: https://github.com/tencent/AngelSlim
19+
- License: Apache License 2.0
20+
- Copyright notice retained in source files:
21+
`Copyright 2025 Tencent Inc. All Rights Reserved.`
22+
- Local usage: quantization utilities under `quantizer/`.
23+
24+
## LightX2V
25+
26+
- Upstream: https://github.com/ModelTC/LightX2V
27+
- Local usage: optional optimized low-precision inference operators used by the
28+
quantized DiT path. This dependency is installed separately and is not
29+
vendored in this repository.
30+
31+
## taehv
32+
33+
- Upstream: https://github.com/madebyollin/taehv
34+
- Local usage: TAeW2.2 lightweight streaming VAE wrapper and related code paths.
35+
36+
## Helios
37+
38+
- Upstream: https://github.com/PKU-YuanGroup/Helios
39+
- Local usage: optimized Triton RoPE and normalization kernels under
40+
`wan/modules/helios_kernels/`.
41+
42+
## Demo Assets
43+
44+
The sample images and pre-generated reference-image cache files under
45+
`web_client/datasets/images/` and `outputs/ref_image_cache/` are included for
46+
running the demo presets. Replace or remove them before redistribution if your
47+
release process requires separate asset provenance records.

configs/default_config.yaml

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,6 @@
1+
vae_type: "taew2_2"
2+
use_fp8_gemm: true
3+
quant_type: "fp8-per-token"
4+
5+
taew2_2_checkpoint: "checkpoints/ABot-World-0-5B-LF/taew2_2.pth"
6+
lightvae_encoder_checkpoint: "checkpoints/ABot-World-0-5B-LF/Wan2.2_VAE.pth"

configs/long_forcing_dmd.yaml

Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
model_type: ci2v
2+
3+
denoising_step_list:
4+
- 1000
5+
- 750
6+
- 500
7+
- 250
8+
warp_denoising_step: true
9+
context_noise: 0
10+
11+
image_or_video_shape:
12+
- 1
13+
- 31
14+
- 48
15+
- 44
16+
- 80
17+
independent_first_frame: true
18+
num_frame_per_block: 3
19+
20+
ref_num_slots: 5
21+
ref_resolution: 512
22+
23+
model_kwargs:
24+
model_name: checkpoints/ABot-World-0-5B-LF
25+
timestep_shift: 5.0
26+
local_attn_size: 21
27+
use_relative_rope: true
28+
downscale_factor_control_adapter: 16
29+
30+
text_encoder_kwargs:
31+
tokenizer_path: checkpoints/ABot-World-0-5B-LF/google/umt5-xxl/
32+
encoder_pth_path: checkpoints/ABot-World-0-5B-LF/models_t5_umt5-xxl-enc-bf16.pth
33+
34+
vae_kwargs:
35+
vae_type: Wan2.2_VAE
36+
pretrained_path: checkpoints/ABot-World-0-5B-LF/Wan2.2_VAE.pth

0 commit comments

Comments
 (0)