Robbyant opens hour-long AI world generation with LingBot-World 2.0

Robbyant, an embodied AI company within Ant Group, has open-sourced LingBot-World 2.0 (Infinity), a generative world model capable of hour-long continuous video generation at 720p and 60 frames per second. This represents a significant step forward from LingBot-World 1.0, which generated stable worlds only for minutes. The new model is engineered to generate, stream, and display content simultaneously, reducing latency and enabling real-time user interaction as the environment unfolds.

The technical architecture centers on a dual-agent mechanism. The Pilot Agent plans and executes character actions, while the Director Agent dynamically introduces new events as the virtual world evolves. The system is built on a causal pretraining approach combined with Robbyant's proprietary Mask of Bidirectional Attention (MoBA) mechanism, specifically designed to reduce the visual degradation that typically occurs during extended AI-generated sequences. According to Robbyant, internal stress testing demonstrated stable visual fidelity throughout hour-long generation sessions without noticeable quality drift.

Users interact with the environment through the Reactor platform, controlling character movement and changing viewpoints in real time using a keyboard. The expanded action set includes attacking, jumping, gliding, casting spells, and shooting arrows. Through text prompts, users can trigger environmental events such as weather changes, day-night cycles, and the introduction of new entities. A notable feature is support for multiple users sharing a persistent virtual environment, enabling collaborative interaction within the same AI-generated world.

Alongside LingBot-World 2.0, Robbyant has also open-sourced LingBot-Video, described as the first open-source video generation foundation model based on a Mixture-of-Experts architecture developed specifically for embodied AI. According to the company, LingBot-Video has been designed for robotics applications, improving inference efficiency, physical realism, action understanding, and task completion as AI-generated video models move beyond content creation toward real-world robotic systems.

LingBot-World 2.0 represents a technical milestone in extending AI-generated environments beyond short clips to sustained, interactive spaces. The core innovation lies in the proprietary Mask of Bidirectional Attention (MoBA) mechanism, which the company designed to combat visual degradation—a well-known failure mode when AI models generate video over long sequences. By combining causal pretraining with this attention architecture, Robbyant has achieved what internal testing confirms: stable visual fidelity throughout hour-long generation without noticeable quality drift.

The dual-agent design (Pilot Agent for character actions, Director Agent for dynamic event injection) transforms the model from a passive video generator into an interactive system. Users no longer simply watch; they shape the world through keyboard input and text prompts, triggering weather changes, day-night cycles, and new entities. The addition of multi-user support in a persistent environment further signals a shift away from isolated simulation toward collaborative virtual worlds.

The simultaneous release of LingBot-Video, an open-source video generation model built on a Mixture-of-Experts architecture and targeting robotics, suggests Robbyant is positioning embodied AI and robotics as the primary use case. Video generation for robotics differs from entertainment content generation—it must capture physical realism, action understanding, and task completion. This pairing positions LingBot-World 2.0 not merely as a creative tool but as infrastructure for training and simulating robotic systems in realistic, interactive environments.

No comments yet. Be the first to share your thoughts!