Primary use stated or inferred cautiously from official documentation.
WALL-WM 2.0 RoboDojo Preview
X Square Robot released this pretrained checkpoint on November 21, 2026 as the starting point for RoboDojo training. It is a preview of WALL-WM 2.0, distinct from the full release, mainly in the number of training iterations. The repository provides weights, configuration and action and state normalization statistics. It is neither a conversational assistant nor a robot policy ready to connect to any machine.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | x-square-robot/wall-wm-2.0-robodojo |
|---|---|
| Model type | Robotics · Multimodal |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 21 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image · Robot state |
| Exit | Video and action representations |
| Context window | Published configuration: three views; text features of length 512 |
| maximum output | 20-dimensional action configuration and 32-step chunks |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Starting point for RoboDojo training |
| Verified platforms | Hugging Face |
Research on world-action models and subsequent adaptation to robot manipulation. The configuration uses Wan2.2-TI2V-5B, and the card identifies the Qwen3.5-9B tokenizer and chat template. It defines three camera views and actions for both arms through relative positions, 6D rotations and grippers. These describe the training artifact; they do not certify compatibility with a particular robot or autonomous execution quality.
The weight file is about 24.1 GB; its size is not a VRAM requirement. Length 512 refers to text features in this configuration, not a universal conversation window. Pin the revision, review dependencies and normalization, and validate in simulation first. Training options do not establish that action inference is enabled or ready without adaptation. The full version and the WALL-OSS policy are different artifacts.
What it can do
Orientation
Research on world-action models and subsequent adaptation to robot manipulation. The configuration uses Wan2.2-TI2V-5B, and the card identifies the Qwen3.5-9B tokenizer and chat template. It defines three camera views and actions for both arms through relative positions, 6D rotations and grippers. These describe the training artifact; they do not certify compatibility with a particular robot or autonomous execution quality.
Context
Published configuration: three views; text features of length 512 · 20-dimensional action configuration and 32-step chunks
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
The card declares Apache 2.0 for this downloadable checkpoint. No commercial endpoint or official tariffs have been verified for this version; numeric prices remain unpublished. Self-hosting consumes compute, memory and operations. The general WALL-WM code explains the architecture but does not by itself establish an identical training recipe for this preview.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
No verified published scores are available for this exact checkpoint. Results from WALL-OSS, the original WALL-WM work or fine-tuned RoboDojo policies do not transfer to these pretrained weights. The 20 action components, 32-step chunks and file size are technical specifications, not measures of robot success. Performance comparisons require the same task, robot, fine-tuning and protocol.
i No verified published results for this preview checkpoint.
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
x-square-robot/wall-wm-2.0-robodojo
Preview weights released; configuration updated on November 22.