Ant Group’s LingBot Technology, in partnership with esteemed academic institutions including the Hong Kong University of Science and Technology, has secured acceptance of its collaborative research paper, titled “Causal World Modeling for Robot Control,” at RSS 2026—a premier academic conference in the field of robotics. This groundbreaking study introduces an innovative causal world modeling framework, which has been successfully translated into an open-source autoregressive video-action world model known as LingBot-VA. This model endows robots with the remarkable ability to anticipate dynamic environmental changes and autonomously generate action commands, enabling them to “observe, assess, and act” in unison, akin to human cognitive processes.

LingBot-VA leverages a sophisticated Mixture-of-Transformers architecture combined with a closed-loop reasoning mechanism, showcasing exceptional performance across both simulated benchmark tests and real-world robotic tasks. Specifically, in the Easy and Hard configurations of the RoboTwin 2.0 dual-arm manipulation tasks, the model achieved average success rates of 92.0% and 91.1%, respectively. In the LIBERO benchmark test, its success rate soared to an impressive 98.5%. During real-world evaluations, LingBot-VA demonstrated a substantial improvement in overall success rates, surpassing industry benchmarks by over 20 percentage points. Furthermore, it required only 50 real demonstration data points for adaptation, highlighting its remarkable data efficiency and generalization capabilities.

Currently, the model weights and code for LingBot-VA are publicly accessible.