Amap CV Lab · robot AI model
ABot-M0.5
China · 2026
Sources retrieved: 2 November 2026 (A, B, C, D)
- Model type
- World action modelSource and note: Model typeC arxiv.org/abs/2607.00678 · Retrieved 2 November 2026Note. Paper title: “ABot-M0.5: Unified Mobility-and-Manipulation World Action Model”. Abstract: “We propose ABot-M0.5, a new WAM built on the insight that mobile manipulation requires alignment at three levels”.
- What it controls
- Arms and hands Whole bodySource and note: What it controlsA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. Technical report: “it jointly models future video latents, frame-level latent actions, and executable robot actions within a unified generative framework”. The paper names the action parts: “heterogeneous action subspaces such as base movement and arm manipulation”.
- Inputs
- Camera LanguageSource and note: InputsA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. Technical report: “Given a language instruction l l and a sequence of multi-view observations”.
- Access
- Open weights (download)Source and note: AccessB huggingface.co/acvlab/ABot-M0.5-RoboCasa365 · Retrieved 2 November 2026Note. Amap CV Lab's code repository, July 2026: “The inference code, pre-trained weights, and evaluation scripts for ABot-M0.5 are now available.” The Hugging Face checkpoint repository is gated with automatic approval, so its files need a login and accepted conditions.
- Base model
- Wan2.2 5BSource and note: Base modelA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. Technical report: “Initialized from the pretrained Wan2.2 5B weights”.
- Licence type
- Free, commercial useSource and note: Licence typeD github.com/amap-cvlab/ABot-Manipulation/blob/main/GETTING_STARTED.md · Retrieved 2 November 2026Note. Code repository: “Licensed under the Apache License, Version 2.0”.
Specifications
9/14 specifications disclosed 5 not stated
3 fields hidden 2 not-stated fields with a note stay ·
How to read the figuresShow guideClose
Not statedThe developer publishes nothing on it
Select a source marker for source details and notes.
Release1/1 disclosed
Released1 July 2026
Source and note: ReleasedC arxiv.org/abs/2607.00678 · Retrieved 2 November 2026Note. [A] “Submitted on 1 Jul 2026”
Input and output3/4 disclosed
Model typeWorld action model
Source and note: Model typeC arxiv.org/abs/2607.00678 · Retrieved 2 November 2026Note. Paper title: “ABot-M0.5: Unified Mobility-and-Manipulation World Action Model”. Abstract: “We propose ABot-M0.5, a new WAM built on the insight that mobile manipulation requires alignment at three levels”.
InputsCamera Language
Source and note: InputsA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. Technical report: “Given a language instruction l l and a sequence of multi-view observations”.
What it controlsArms and hands Whole body
Source and note: What it controlsA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. Technical report: “it jointly models future video latents, frame-level latent actions, and executable robot actions within a unified generative framework”. The paper names the action parts: “heterogeneous action subspaces such as base movement and arm manipulation”.
Degrees of freedom controlledNot stated
Size and speed0/2 disclosed
ParametersNot stated
Source and note: ParametersA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. The report names its starting weights, not a total count: “Initialized from the pretrained Wan2.2 5B weights”.
Control rateNot stated
Where it runs and what it is built on1/2 disclosed
On robot or cloudNot stated
Base modelWan2.2 5B
Source and note: Base modelA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. Technical report: “Initialized from the pretrained Wan2.2 5B weights”.
Openness3/3 disclosed
AccessOpen weights (download)
Source and note: AccessB huggingface.co/acvlab/ABot-M0.5-RoboCasa365 · Retrieved 2 November 2026Note. Amap CV Lab's code repository, July 2026: “The inference code, pre-trained weights, and evaluation scripts for ABot-M0.5 are now available.” The Hugging Face checkpoint repository is gated with automatic approval, so its files need a login and accepted conditions.
Licence documentApache-2.0
Source and note: Licence documentB huggingface.co/acvlab/ABot-M0.5-RoboCasa365 · Retrieved 2 November 2026Note. Hugging Face checkpoint page: “License: apache-2.0”. Code repository: “Licensed under the Apache License, Version 2.0”.
Licence typeFree, commercial use
Source and note: Licence typeD github.com/amap-cvlab/ABot-Manipulation/blob/main/GETTING_STARTED.md · Retrieved 2 November 2026Note. Code repository: “Licensed under the Apache License, Version 2.0”.
The developer’s own claims1/2 disclosed
Learns new tasksNeeds retraining
Source and note: Learns new tasksA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. The report's real-world tests: “Experiments were conducted on an Agilex Piper robotic arm (6-DoF, single arm)”; “only 50 real-world demonstrations were collected for each task”.
Works across robotsNot stated
Source and note: Works across robotsA arxiv.org/html/2607.00678 · Retrieved 2 November 2026Note. The report says: “making the framework adaptable to diverse robot embodiments”. Its real-world tests use one robot arm: “Experiments were conducted on an Agilex Piper robotic arm (6-DoF, single arm)”.Yes 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
Amap CV Lab published the ABot-M0.5 technical report on arXiv on 1 July 2026.
The ABot-M0.5 code repository describes the model as built on the Wan2.2 video diffusion backbone.
Sources
One line per source, not per figure: a source has one retrieval date.
- Aarxiv.org/html/2607.00678Retrieved 2 November 2026
- Bhuggingface.co/acvlab/ABot-M0.5-RoboCasa365Retrieved 2 November 2026
- Carxiv.org/abs/2607.00678Retrieved 2 November 2026
- Dgithub.com/amap-cvlab/ABot-Manipulation/blob/main/GETTING_STARTED.mdRetrieved 2 November 2026