On July 10, Ant Lingbo released the industry's first embodiment-native world action model, LingBot-VA2.0. The release of this model marks a key transformation in robot foundation models, from "building based on digital world models" to "native design for the physical world." It represents a critical path choice in the development of embodied intelligence: the robot "brain" no longer relies on the "grafting" of digital world model capabilities but is instead designed natively, starting from fundamental requirements such as dynamic modeling, causal prediction, and real-time execution in interactions with the environment.

Thanks to its embodiment-native architecture, LingBot-VA2.0 has demonstrated outstanding execution speed and generalization ability in real machine tests. Take the following video as an example; the robot can complete multi-round random sparring with humans without relying on any external camera equipment.

image.png

This year, how to integrate world models with embodied intelligence has been a focal point for all parties. Starting from the end goal and focusing on the "control execution" needs of the physical world, it is necessary to have continuous causal prediction capabilities. Robots face a continuously changing real world; they not only need to react to current situations but also understand what environmental changes an action will cause and decide on the next action accordingly. Most mainstream approaches in the industry rely on video generation models designed for digital content creation and then adapt them to robot control tasks through fine-tuning.

However, content creation and robot control have different starting points. Content creation focuses more on image quality and creativity, while robot control emphasizes execution efficiency and the rationality of predictions. These differences lead to distinct capability emphases between video models in the digital world and video action models in the physical world from the design stage. Forcing "fine-tuning" to adapt the former to the latter can result in side effects like knowledge forgetting and reduced generalization ability.

LingBot-VA2.0 directly addresses the problem and explores a more difficult path—pre-training from scratch based on an autoregressive architecture, building a native base model through four core designs.

First, the model introduces a semantic visual-action tokenizer as a new visual encoder, adding alignment of semantic and action information during the visual compression process. This makes it easier for the model to convert "understanding instructions" into "performing actions" in subsequent training, which helps in instruction following and improves action accuracy. Second, the model adopts a strict causal pre-training paradigm, allowing the model to use an autoregressive architecture from the start of training, ensuring that visual prediction and action generation strictly follow a unidirectional time sequence. Third, it introduces the MoE architecture, effectively expanding the model capacity without sacrificing inference efficiency, achieving a balance between performance and efficiency. Finally, by enhancing the asynchronous inference mechanism, real-time closed-loop control is achieved. While the robot performs actions, it predicts future states and continuously corrects the next decision using the latest real observations. Based on these designs, LingBot-VA2.0 provides a response of 150Hz real-time inference efficiency per card, addressing the common issue of low execution efficiency in embodied world models in the industry.

From the perspective of "working," robots need to "see more clearly," "think more clearly," and "work more efficiently." This week, Ant Lingbo has continuously released and open-sourced multiple models, including LingBot-Vision and LingBot-Depth2.0 for spatial perception, LingBot-VLA2.0 for "one brain for multiple machines," LingBot-World2.0 for real-time interaction, and LingBot-Video for higher reasoning efficiency. These models represent Ant Lingbo's continuous exploration of specialized capabilities required for embodiment-native intelligence. As a culmination, LingBot-VA2.0 plays the role of a closing work and officially ushers in a new phase of embodiment-native intelligence.

Zhu Xing, CEO of Ant Lingbo, stated that on one hand, Lingbo will continue to explore the new limits of embodied intelligence, and on the other hand, accelerate the construction of an open technical and scenario ecosystem, helping robots to rapidly enter industrial scenarios.

It is reported that Ant Lingbo will comprehensively demonstrate the capabilities of the full-stack brain 2.0 during the 2026 World Artificial Intelligence Conference (WAIC) from July 17 to 20. Visitors can experience it on-site at the Shanghai World Expo Exhibition Hall (H3-B302 and H1-C701 booths).