Robbyant, an embodied AI company within China’s Ant Group, announced LingBot-VA 2.0, describing it as the industry’s first embodied native video action world model for robotics.
The model is built from the ground up for physical-world tasks, rather than adapting a video generation model originally designed for digital content creation.
LingBot-VA 2.0 uses an autoregressive architecture to predict how the robot’s actions will change the environment and determine the next action based on those causal relationships.
According to the company, this approach improves physical precision, execution efficiency, and versatility for real-world robotic applications.
Redefining robot learning
Robbyant describes LingBot-VA 2.0 as the industry’s first embody-native video action world model built specifically for robotics, rather than an applied digital content generation system.
This release represents a shift in the robotics foundation model by designing AI natively for the physical world. Unlike traditional approaches that fine-tune video generation models for robot control, LingBot-VA 2.0 was pre-trained from scratch using an autoregressive architecture focused on dynamic world modeling, causal prediction, and real-time execution. The company says this allows the model to predict how the robot’s actions will change its surroundings, and then choose its next action based on those predictions.
Most existing body-based AI systems rely on video models originally developed to generate digital content. These models are effective at creating realistic visuals, but prioritize image quality and creativity over physical accuracy and execution speed. Robyant says adapting these to robotics reduces generalizability and often limits real-world performance.
LingBot-VA 2.0 addresses these challenges through four architectural innovations. A semantic visual action tokenizer jointly compresses visual and action information, allowing the model to properly translate instructions into robot movements. A strict causal pre-training strategy ensures that predictions follow the correct temporal sequence, while a mixture of experts (MoE) architecture improves the model’s power without sacrificing inference efficiency. Enhanced asynchronous reasoning mechanisms allow robots to predict future states while performing actions and continuously update decisions using real-world observations.
The company says these advances enable real-time closed-loop control at 150Hz on a single GPU. The model can adapt to new manipulation tasks in just 20 demonstrations through in-context learning, eliminating the need to update parameters.
predictive robot intelligence
LingBot-VA integrates future video prediction and policy learning within a single autoregressive framework to jointly learn visual dynamics and robot actions. It is pre-trained on a large-scale robot video action dataset before being fine-tuned for downstream tasks.
During operation, the system first predicts future visual states from current observations and verbal instructions. An inverse dynamics model then translates these predictions into executable robot behavior, while real-world observations continuously replace the predicted states to keep the control loop grounded in reality.
Robyant demonstrated the model through long, precise operational tasks such as preparing breakfast, unpacking deliveries, inserting tubes, picking up screws, folding clothes, and opening drawers. The company also reported superior results over existing methods across multiple task settings on the RoboTwin 2.0 and LIBERO simulation benchmarks.
The company also highlighted LingBot-VA’s ability to retain long-term memory, allowing the robot to distinguish between visually identical but contextually different situations and accurately perform multi-step tasks that require counting, sequencing, and repetitive actions.
“Robby Ant will continue to explore new boundaries in embodied intelligence while accelerating the development of open technology and application ecosystems to accelerate the adoption of robots in industrial and real-world scenarios,” said Zhu Xing, CEO of Robby Ant, in a statement.
