Ant Group’s robotics unit Robbyant has published the LingBot-VA 2.0 technical report, introducing what it calls the first Physical AI video-action foundation model built from scratch for embodiment—not fine-tuned from a video generator.
Built for Real-Time Control
Unlike previous approaches that adapt video generation models for robotics, LingBot-VA 2.0 was designed from the ground up for physical interaction. The model achieves 225 Hz asynchronous control, enabling real-time feedback loops with robotic systems.
The key technical innovations include:
- Foresight Reasoning: Predicts future states ahead of execution, allowing the robot to plan several steps ahead
- Re-grounding on Every Observation: Continuously updates its internal model based on real sensory input
- Causal DiT Architecture: A causal Diffusion Transformer specifically designed for action prediction
- Sparse-MoE Video Stream: Mixture of Experts design for efficient video processing
How It Differs from VLA Approaches
Vision-Language-Action (VLA) models like Google’s RT-2 and DeepMind’s RT-3 typically adapt large language or vision models for robot control. LingBot-VA 2.0 takes a different path by building a native action model that understands the physics of movement from training data.
The model was pretrained on approximately 60,000 hours of robot trajectory data, covering 20 different robot configurations and 10,000 hours of egocentric human video. It maps every embodiment into a 55-dimensional canonical action space, supporting arms, dexterous hands, waists, heads, and mobile bases.
Industry Implications
This release signals a shift in the Physical AI landscape. While companies like Tesla (Optimus), Boston Dynamics, and Figure AI focus on humanoid robots, Ant Group’s approach emphasizes general-purpose manipulation across different robot forms.
The 225 Hz control rate is particularly notable—most commercial robot systems operate at 50-100 Hz. This higher frequency enables smoother, more responsive control for delicate tasks like object manipulation and tool use.
Robbyant has released the technical report with performance benchmarks, though deployment code and quantitative evaluations are still limited. The model is available for research purposes under an Apache 2.0 license.