Robbyant launches LingBot-VA 2.0 embodied-native world-action model

Robbyant, an embodied AI company within Ant Group, today announced the release of LingBot-VA 2.0, an embodied-native world-action model.

The company said that this release marks a key transition in robotics foundation models, shifting from repurposing digital world models to designing them natively for the physical world. Instead of relying on fine-tuned digital content generation models, Robbyant said that LingBot-VA 2.0 is built from scratch to meet the original demands of dynamic modeling, causal prediction and real-time execution in physical environments.

A new era of robotics world action models

Robbyant said that integrating world models with embodied AI has been one of the major focuses of the AI industry. However, most mainstream approaches rely on video generation models designed for digital content, which are then fine-tuned for robot control. Because content creation prioritizes visual quality and creativity, while robot control requires execution efficiency and physical accuracy, the company said that this forced adaptation often leads to knowledge forgetting and reduced generalization.

��Robots that think ahead and act in real time.

LingBot-VA 2.0 — the first embodied-native foundation model. Not fine-tuned from a video generator. Built from scratch for the physical world.

✅ 93.6% success on bimanual tasks
⚡ 150 Hz single-GPU inference
�� 20 demos to… pic.twitter.com/kV8ENPngL8

— Robbyant (@robbyant_brain) July 9, 2026

Robbyant said that LingBot-VA 2.0 takes a different approach. By pre-training from scratch using an autoregressive architecture, the model is designed to understand how an action will change the environment and to decide the next step based on that causal prediction.

Robbyant said that to achieve this new approach, LingBot-VA 2.0 is built on four core designs:

• Semantic visual-action tokenizer: A new visual encoder that aligns semantic and action information during visual compression, helping the model translate "understanding instructions" into "completing actions" more effectively.
• Strict causal pre-training: The model uses an autoregressive architecture from the beginning, ensuring that visual predictions and action generation strictly follow a one-way time sequence.
• Mixture of Experts (MoE): This architecture expands model capacity without sacrificing inference efficiency, balancing performance and speed.
• Enhanced asynchronous inference: This mechanism enables real-time closed-loop control, allowing robots to predict future states while executing actions and continuously corrects its next decisions using the latest real-world observations.

“Robbyant will continue to explore new limits in embodied intelligence while accelerating the development of an open technology and application ecosystem to expedite robot deployment in industrial and real-world scenarios,” said Zhu Xing, CEO of Robbyant.

Robbyant said that these designs solve the common industry challenge of low execution efficiency in embodied world models, delivering a real-time inference speed of 150 Hz on a single GPU. Furthermore, the model can generalize to new tasks using as few as 20 demonstrations through in-context learning without parameter updates. 

Robbyant said that LingBot-VA 2.0 serves as the capstone of its recent launch week, which introduced six models.

The company said that when combined, these models form a complete, embodied-native full-stack for perception, world simulation and action. The launches include: 

• LingBot-Depth 2.0

• LingBot-Vision
• LingBot-VLA 2.0
• LingBot-World 2.0
• LingBot-Video
• LingBot-VA 2.0