Embodied AI is rapidly emerging as the next frontier of artificial intelligence, shifting the focus from systems that understand the digital world to robots capable of perceiving, reasoning, and acting in complex physical environments. At the heart of this evolution are embodied-native world models that integrate perception, prediction, and action into a unified intelligence framework, enabling robots to learn faster, adapt with minimal data, and perform real-time decision-making. As industries increasingly look to intelligent robots for applications ranging from manufacturing and logistics to healthcare and household assistance, advances in foundation models are poised to redefine the future of robotics. In this interview with AI Spectrum, Yujun Shen, Chief Scientist at Robbyant, discusses the company's embodied-native video-action world model, LingBot-VA 2.0, and explains how causal reasoning, multimodal intelligence, and real-time world modelling are laying the foundation for the next generation of autonomous robots.

Robbyant describes LingBot-VA 2.0 as the industry’s first embodied-native video-action world model. How does this architecture fundamentally differ from conventional robotics AI models that are adapted from digital video generation systems?

The key difference is that LingBot-VA 2.0 is designed as an embodied-native world model from the very beginning, rather than adapting a general-purpose video generation model for robotics.

Most existing video generation models are trained with bidirectional objectives that focus on producing visually plausible videos. While this works well for digital content generation, it does not naturally respect the causal structure required for robot control.

LingBot-VA 2.0 instead adopts a causal autoregressive video-action architecture. The model predicts future observations and robot actions strictly in temporal order, where every action influences future observations and future information is never available during decision-making. This causal formulation better reflects how robots interact with the physical world.

Rather than simply generating videos, the model learns action-conditioned world dynamics during large-scale embodied pretraining. This enables more reliable closed-loop control and makes the model fundamentally better suited for robotics than approaches adapted from digital video generation.

One of the biggest challenges in embodied AI is enabling robots to understand and interact with dynamic physical environments. How does LingBot-VA 2.0 improve causal reasoning, real-time decision-making, and task generalisation compared with existing world models?

We improve all three aspects through both model design and system optimisation.

For causal reasoning, LingBot-VA 2.0 is causally pretrained from the beginning. Instead of learning to reconstruct videos, the model learns how future observations evolve as a consequence of robot actions, allowing it to better capture the causal relationship between actions and environmental changes.

For real-time decision-making, we redesigned the entire inference stack and optimised both the model architecture and deployment pipeline. Compared with the previous generation, inference is approximately four to five times faster, enabling real-time closed-loop control at up to 150 Hz.

For task generalisation, large-scale embodied pretraining significantly reduces the amount of downstream data required. In many practical scenarios, only 10 to 20 demonstrations are sufficient to fine-tune a high-quality policy for a new manipulation task.

These improvements are made possible through embodied-native pretraining, architectural innovations, backend acceleration, and high-quality robot data curation.

LingBot-VA 2.0 reportedly achieves real-time inference at 150 Hz and can generalise to new tasks using as few as 20 demonstrations. What technological innovations have enabled these performance gains, and what practical advantages do they offer for commercial robotics deployments?

Several innovations contribute to these performance gains.

First, we introduce Foresight Reasoning, which enables the model to explicitly predict future world evolution and anticipate the consequences of robot actions before they are executed. This allows robots to make more informed decisions in dynamic environments.

Second, we redesigned the entire deployment stack, including an optimised one-step inference pipeline and extensive backend acceleration, enabling real-time inference at up to 150 Hz.

Another important factor is data quality. Through large-scale embodied pretraining and carefully curated robot datasets, the model learns robust representations that allow it to adapt efficiently from only a small number of demonstrations.

For commercial robotics, these improvements translate into smoother robot behaviours, lower deployment latency, reduced data collection costs, and much faster adaptation to new tasks and new environments.

Embodied AI is increasingly being viewed as the next frontier of artificial intelligence. Which industries or use cases do you believe will be the earliest beneficiaries of embodied-native world models, and how do you see adoption evolving over the next five years?

We believe household service robots and industrial manipulation will be among the earliest beneficiaries of embodied-native world models.

These environments involve long-horizon tasks, dynamic interactions, and highly diverse objects, making them particularly suitable for world models that can understand and predict physical interactions.

Over the next five years, we expect robots to evolve from executing predefined skills toward continuously learning new tasks with minimal demonstrations. Foundation world models will gradually become a common intelligence backbone across different robotic platforms, enabling faster deployment, lower development costs, and broader adoption of embodied AI.

As embodied AI systems become more autonomous, ensuring safety and reliability will be critical. What safeguards and validation processes has Robbyant incorporated into LingBot-VA 2.0 to enable dependable operation in real-world environments?

Safety is paramount for robotics. At Robbyant, we place immense importance on the safety and reliability of embodied AI. We are committed not only to maintaining rigorous, rule-based safety restrictions, but also to exploring how to enable the "embodied AI brain" to achieve intrinsic safety.

Looking ahead, what is Robbyant's long-term vision for embodied AI? How do you see world models, multimodal reasoning, and foundation models shaping the future of intelligent robots, particularly within the broader AI ecosystem being developed by Ant Group?

Guided by the core philosophy of "embodied-native design," Robbyant is dedicated to building an embodied intelligence platform tailored for real-life use cases. Our long-term vision is to bring intelligent services into people's daily lives. Ultimately, we envision robotic companions and caregivers that truly understand and enhance people’s everyday lives, delivering reliable intelligent services across key use cases such as elderly care, medical assistance, and household tasks.