X Square Robot · robot AI model
HOST
China · 2026
Sources retrieved: 2 November 2026 (A, B, C, D)
- Model type
- Vision-language-action (VLA)Source and note: Model typeB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Hugging Face model card tag: “vision-language-action”. Model card: “a single autoregressive diffusion model with dual video/action experts (a Mixture-of-Transformers architecture)”.
- What it controls
- Arms and handsSource and note: What it controlsA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. Paper: “Each arm operates in a 10-dimensional action space comprising Cartesian position, a 6D rotation representation, and gripper aperture, yielding a 20-dimensional bimanual action space”.
- Inputs
- Camera Language Robot stateSource and note: InputsA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. Paper: “The framework accepts either a language instruction, a video demonstration, or both as task specification at inference time.” “Three RGB cameras observe the workspace”; “It also comprises the proprioceptive state, projected via a learnable linear layer”.
- Access
- Open weights (download)Source and note: AccessB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Hugging Face model card: “Self-Grounded Prediction (policy_training/) — released checkpoint”.
- Base model
- Wan2.2-TI2V-5BSource and note: Base modelA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. Paper, video expert: “initialized from Wan2.2-TI2V-5B”.
- Licence type
- Free, commercial useSource and note: Licence typeB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Developer wording: “license: mit”.
Specifications
9/14 specifications disclosed 5 not stated
1 field hidden 4 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
Released22 July 2026
Source and note: ReleasedD arxiv.org/abs/2607.20033 · Retrieved 2 November 2026Note. [A] “[v1] Wed, 22 Jul 2026 11:20:05 UTC”
Input and output3/4 disclosed
Model typeVision-language-action (VLA)
Source and note: Model typeB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Hugging Face model card tag: “vision-language-action”. Model card: “a single autoregressive diffusion model with dual video/action experts (a Mixture-of-Transformers architecture)”.
InputsCamera Language Robot state
Source and note: InputsA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. Paper: “The framework accepts either a language instruction, a video demonstration, or both as task specification at inference time.” “Three RGB cameras observe the workspace”; “It also comprises the proprioceptive state, projected via a learnable linear layer”.
What it controlsArms and hands
Source and note: What it controlsA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. Paper: “Each arm operates in a 10-dimensional action space comprising Cartesian position, a 6D rotation representation, and gripper aperture, yielding a 20-dimensional bimanual action space”.
Degrees of freedom controlledNot stated
Source and note: Degrees of freedom controlledA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. The paper gives a size for the action space, not a number of controlled degrees of freedom: “Each arm operates in a 10-dimensional action space comprising Cartesian position, a 6D rotation representation, and gripper aperture, yielding a 20-dimensional bimanual action space”. Robot: “All experiments are conducted on a bimanual manipulation platform consisting of two ARX R5 six-axis robotic arms”.
Size and speed0/2 disclosed
ParametersNot stated
Source and note: ParametersA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. The paper gives layer sizes for its two experts, not a total parameter count.
Control rateNot stated
Source and note: Control rateA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. The paper gives the data frame rate, not a control rate: “The base action frame interval is 15 frames at 32 Hz”.
Where it runs and what it is built on1/2 disclosed
On robot or cloudNot stated
Base modelWan2.2-TI2V-5B
Source and note: Base modelA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. Paper, video expert: “initialized from Wan2.2-TI2V-5B”.
Openness3/3 disclosed
AccessOpen weights (download)
Source and note: AccessB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Hugging Face model card: “Self-Grounded Prediction (policy_training/) — released checkpoint”.
Licence documentMIT
Source and note: Licence documentB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Model card header: “license: mit”. GitHub: “This snapshot does not yet include a repository-wide license file.”
Licence typeFree, commercial use
Source and note: Licence typeB huggingface.co/Guangyan/HOST · Retrieved 2 November 2026Note. Developer wording: “license: mit”.
The developer’s own claims1/2 disclosed
Learns new tasksFrom a few examples
Source and note: Learns new tasksC github.com/CGuangyan-BIT/HOST · Retrieved 2 November 2026Note. Developer claim: “a framework that enables a robot to acquire skills in seconds from a single human video while retaining previously mastered skills”. “Because skill acquisition does not modify the policy weights, previously mastered skills are retained.”
Works across robotsNot stated
Source and note: Works across robotsA arxiv.org/html/2607.20033v1 · Retrieved 2 November 2026Note. The paper's robot tests use one platform: “All experiments are conducted on a bimanual manipulation platform consisting of two ARX R5 six-axis robotic arms”.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
The HOST project page names three institutions: Beijing Institute of Technology, X SQUARE ROBOT and Tsinghua University.
The paper states that part of the work was done at X SQUARE ROBOT.
The released checkpoint was trained on robot demonstrations only.
Sources
One line per source, not per figure: a source has one retrieval date.
- Aarxiv.org/html/2607.20033v1Retrieved 2 November 2026
- Bhuggingface.co/Guangyan/HOSTRetrieved 2 November 2026
- Cgithub.com/CGuangyan-BIT/HOSTRetrieved 2 November 2026
- Darxiv.org/abs/2607.20033Retrieved 2 November 2026