RobbyantChina2026
LingBot-VA
An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.
Robotsnot stated
Key figures as stated by the developer
5 of 6 stated · 1 gap- Model type not stated
-
Controls A
arms and hands -
Inputs A
cameralanguage - Access B open download
- Parameters A 5,300,000,000
- Base model A Wan2.2-5B (video stream)
How to read the figures
- 14kgA30.9lbA
- The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
- 74.5lb
- Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.
Versions and notes in the key figures (6)
Model type
Note The developer calls it something else and does not label it vision-language-action, world action model or embodied reasoning: "an autoregressive diffusion world model that addresses these limitations through a unified video-action framework".
Controls
"We use a dual-arm representation where each robotic arm is characterized by both end-effector pose (EEF) and joint angles."
Inputs
The model is conditioned on video ("Given a conditioning signal c c (text prompt or initial image)") and on text: "Task instructions are encoded using a frozen T5 text encoder".
Access
The news section of the GitHub README says "Weights and code for shared backbone released!"; the checkpoints are on Hugging Face and ModelScope.
Parameters
"resulting in approximately 350M additional parameters and a total model size of 5.3B parameters".
Base model
"a video stream initialized from Wan2.2-5B".
Full specifications
7 of 12 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Input and output | ||
| Model type | not stated | Note The developer calls it something else and does not label it vision-language-action, world action model or embodied reasoning: "an autoregressive diffusion world model that addresses these limitations through a unified video-action framework". A |
| Inputs | camera language | A The model is conditioned on video ("Given a conditioning signal c c (text prompt or initial image)") and on text: "Task instructions are encoded using a frozen T5 text encoder". |
| Controls | arms and hands | A "We use a dual-arm representation where each robotic arm is characterized by both end-effector pose (EEF) and joint angles." |
| Degrees of freedom controlled | not stated | |
| Size and speed | ||
| Parameters | 5,300,000,000 | A "resulting in approximately 350M additional parameters and a total model size of 5.3B parameters". |
| Control rate | not stated | Note In the RoboTwin 2.0 simulation setup: "We downsample the original 50 Hz video to 12.5 Hz while maintaining the action frequency at 50Hz." A |
| Where it runs and what it is built on | ||
| Runs where | not stated | Note The README says the model "supports distributed inference on GPUs, clusters, and other devices". B |
| Base model | Wan2.2-5B (video stream) | A "a video stream initialized from Wan2.2-5B". |
| Openness | ||
| Access | open download | B The news section of the GitHub README says "Weights and code for shared backbone released!"; the checkpoints are on Hugging Face and ModelScope. |
| License | Apache License 2.0 | C "This project is released under the Apache License 2.0." |
| The developer’s own claims | ||
| Learns new tasks | needs retraining | A Developer claim: "post-training with as few as 50 demonstrations is sufficient for effective deployment". |
| Works across robots | not stated | Note The paper says "To achieve cross-embodiment generalization, we define a universal action interface to adapt to different datasets." It also says "adapting to novel robot platforms requires a small amount of task-specific data". A 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
- Robbyant describes LingBot-VA as an autoregressive diffusion framework that learns frame prediction and policy execution simultaneously.
- The model interleaves video and action tokens in one sequence: a dual-stream mixture-of-transformers design.
- Robbyant released the weights and code for the shared-backbone version on 29 January 2026.
Sources
-
A
Manufacturer product page. arxiv.
org/ · Retrievedhtml/ 2601. 21998 -
B
github.
com/ · RetrievedRobbyant/ lingbot-va -
C
huggingface.
co/ · Retrievedrobbyant/ lingbot-va-base
One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.