Amap CV Lab · robot AI model
ABot-PhysWorld
China · 2026
Sources retrieved: 19 November 2026 (A, B)
- Model type
- World modelSource and note: Model typeA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “a physically consistent, action-controllable video world model for robotic manipulation, built on a 14-billion-parameter Diffusion Transformer.”
- Inputs
- Camera LanguageSource and note: InputsA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README inference supports image and text prompt conditioning.
- Access
- Open weights (download)Source and note: AccessA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “The model checkpoint is auto-downloaded from [ModelScope](https://www.modelscope.cn/models/amap_cvlab/Abot-PhysWorld) on first run.”
- Parameters
- 14BSource and note: ParametersA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “a physically consistent, action-controllable video world model for robotic manipulation, built on a 14-billion-parameter Diffusion Transformer.”
- Base model
- Wan2.1-I2V-14B-480PSource and note: Base modelA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “We provide full-parameter SFT training scripts to fine-tune Wan2.1-I2V-14B-480P on your own robot manipulation datasets.”

Specifications
7 of 14 stated 7 not stated
5 fields hidden 2 not-stated fields with a note stay ·
How to read the figuresShow guideClose
- not statedNo source
- 0Reported as zero
- ASource
Select a source marker for source details and notes.
Release1/1 disclosed
Released24 March 2026
Source and note: ReleasedB arxiv.org/abs/2603.23376 · Retrieved 19 November 2026Note. arXiv abstract page: “Submitted on 24 Mar 2026”
Input and output2/4 disclosed
Model typeWorld model
Source and note: Model typeA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “a physically consistent, action-controllable video world model for robotic manipulation, built on a 14-billion-parameter Diffusion Transformer.”
InputsCamera Language
Source and note: InputsA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README inference supports image and text prompt conditioning.
What it controlsNot stated
Source and note: What it controlsA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. Generates predicted robot manipulation video frames rather than direct robot control actions.
Degrees of freedom controlledNot stated
Size and speed1/2 disclosed
Parameters14B
Source and note: ParametersA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “a physically consistent, action-controllable video world model for robotic manipulation, built on a 14-billion-parameter Diffusion Transformer.”
Control rateNot stated
Where it runs and what it is built on1/2 disclosed
On robot or cloudNot stated
Base modelWan2.1-I2V-14B-480P
Source and note: Base modelA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “We provide full-parameter SFT training scripts to fine-tune Wan2.1-I2V-14B-480P on your own robot manipulation datasets.”
Openness1/3 disclosed
AccessOpen weights (download)
Source and note: AccessA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. GitHub README: “The model checkpoint is auto-downloaded from [ModelScope](https://www.modelscope.cn/models/amap_cvlab/Abot-PhysWorld) on first run.”
Licence documentNot stated
Licence typeNot stated
The developer’s own claims1/2 disclosed
Learns new tasksWithout examples
Source and note: Learns new tasksA raw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.md · Retrieved 19 November 2026Note. Developer claim, GitHub README: “It integrates physics-aware training, memory-efficient preference optimization, and precise spatial action injection to generate realistic and physically plausible robot-object interactions — even in zero-shot settings.”
Works across robotsNot stated
Source and note: Works across robotsYes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.
Notes
Developed by AMAP CV Lab and submitted to arXiv on 24 March 2026. Inference code and ModelScope checkpoints are publicly available.
Sources
One line per source, not per figure: a source has one retrieval date.
- Araw.githubusercontent.com/amap-cvlab/ABot-PhysWorld/main/README.mdMain source. Retrieved 19 November 2026
- Barxiv.org/abs/2603.23376Retrieved 19 November 2026


