The Ant Group subsidiary Robbyant has released LingBot-VA 2.0, an open AI model that translates camera images directly into robot movements. On the simulation benchmark RoboTwin 2.0, the system achieves a success rate of 93.6 percent and operates up to six and a half times faster than the previous version.
Causal architecture connects video and action
According to a paper published on July 9 in arXiv, LingBot-VA 2.0 is the first model in the industry that was natively developed for the physical world, rather than retrofitting an existing video generation model. The system combines a semantic tokenizer that maps world states and actions in a common space with a strictly causal pre-training architecture. This prevents the model from using information from the future, a problem of classical bidirectional video models. Overall, the model comprises 15.3 billion parameters, of which only about 2.5 billion are activated thanks to a mixture-of-experts architecture with 128 experts and top-8 routing per processing step. This efficiency reduces computational effort without limiting the model’s capacity. An additional feature called Foresight Reasoning allows the system to predict future states while the robot is still executing the current action; real sensor data continuously corrects the prediction. The model was trained with five different goals simultaneously, ranging from text-to-image to text-to-video to human-robot co-training, with the weighting gradually shifting from coarse to fine tasks during training.
Benchmark shows significant lead over rivals
The robotics benchmark RoboTwin 2.0 includes 50 manipulation tasks with a total of 2,500 clean and 25,000 randomly varied demonstrations. According to a technical analysis by MarkTechPost, LingBot-VA 2.0 achieves a success rate of 93.6 percent on this benchmark. The comparison model Motus reaches 87.9 percent, the system π0.5 achieves 79.8 percent, and X-VLA scores 72.9 percent. In addition to accuracy, speed also improves significantly: latency decreases from 927 to 142 milliseconds, and the asynchronous control frequency increases from 35 to up to 225 Hertz. This allows the model to make movement decisions significantly more frequently per second than previous versions, enabling smoother responses to changing environments. For new tasks, the system reportedly requires only ten to twenty demonstrations to adapt through in-context learning, without needing to retrain the model parameters. These figures come from simulations and are independently unverified; how the model performs on real robots in uncontrolled environments remains open.
Robbyant builds six-part model stack for robots
According to the company, LingBot-VA 2.0 is part of a six-part model stack for embodied AI. This includes the perception models LingBot-Vision and LingBot-Depth 2.0, the world model LingBot-World 2.0, the video model LingBot-Video, and the control model LingBot-VLA 2.0, which was updated in early July. Together, the stack covers the entire chain from depth and object recognition to simulating future world states to the concrete movement control of a robot. Robbyant operates as an embodied AI unit within the Ant Group, which is primarily known for the payment service Alipay and has recently released new models in this series on a weekly basis. As reported by IT Tech News, the technology is aimed at household robots, applications in elderly care, and medical assistance systems. The open access via GitHub is intended to allow developers to transfer the model to their own robot platforms without having to pay licensing fees to a single provider. This positions Robbyant among a growing number of providers that publish open foundational models specifically for robotics, rather than retrofitting existing language or image models for the physical world.
It will be crucial to see whether the success rate from the RoboTwin simulation can be transferred to real hardware, where sensor noise and mechanical play typically increase the error rate of AI controls. It also remains to be seen which robot manufacturers will integrate the open model into their own products and how Robbyant’s six-part portfolio will compete against established providers from the USA and Europe.
