LingBot-VA 2.0 - A native world action model launched by Ant Lingbo.

What is LingBot-VA 2.0?

LingBot-VA 2.0 is the industry's first embodied native world action model launched by AntLingbo. Based on an autoregressive architecture, it was pre-trained from scratch, enabling the robot to have the general control capability of "inferring and acting simultaneously". The model adopts a causal DiT+sparse MoE architecture with a total of 15.3B parameters and only 2.5B activations during inference. It achieves a 93.6% success rate in dual-arm tasks on the RoboTwin 2.0 benchmark, and single-GPU inference reaches 150Hz.

Main functions of LingBot-VA 2.0

• Visual-Motion Joint PredictionSimultaneously generate predictions of the future world state and the next robot action to achieve "predictive control".
• Long-term mission planningThe high-dimensional Planner breaks down complex goals into structured subtasks, supporting parallel execution by both arms and state memory.
• Real-time closed-loop controlForesight Reasoning is an asynchronous reasoning mechanism that pre-computes the next step while performing the current action, eliminating serial latency.
• Cross-domain generalizationIt maintains a success rate of over 93% in both clean and domain randomized scenarios, adapting to changes in real-world environments.
• Precise operationSupports high-precision force control gripping of thin, fragile objects such as potato chips.

Technical Principles of LingBot-VA 2.0

• Semantic Vision-Action SegmenterUnlike traditional VAEs that only pursue pixel reconstruction, the word segmenter forces the alignment of semantic and action information in visual compression. It implicitly extracts action supervision from continuous frames through inverse dynamics and forward dynamics models, enabling unlabeled network videos to provide training signals.
• Causal pre-training paradigmAbandoning post-processing modifications to the bidirectional video generation model, the model is trained from day one using an autoregressive causal architecture, ensuring that the model learns strictly according to the timeline that can only see the past, naturally matching the unidirectional physical reality of the robot's closed-loop control.
• Sparse MoE architectureThe video backbone uses Mixture-of-Experts, with a total of about 13B parameters, but only 1.9B are activated for inference. Combined with acceleration technologies such as consistent distillation and low-precision compilation, the inference latency is reduced from 965ms to 142ms/chunk.
• Foresight Reasoning: Asynchronous ReasoningThe design predicts, executes, and corrects the closed loop. When the robot performs the current action, the model predicts the next step in parallel and recalibrates when the actual observation is returned, avoiding physical drift caused by pure guesswork. The asynchronous control frequency is increased to 225Hz.

How to use LingBot-VA 2.0

• Visit the official websiteVisit https://technology.robbyant.com/lingbot-va-v2 to learn about the model architecture, demo videos, and technical details.
• Get codeVisit the GitHub repository to download the open-source code and technical reports to familiarize yourself with the implementation details of causal DiT, sparse MoE, and Foresight Reasoning.
• Environment DeploymentConfigure the inference environment on a server or workstation equipped with a high-performance GPU, install dependencies, and load model weights.
• Hardware accessConnect the model to the robot body, link the camera observation stream to the robotic arm/gripper actuator, and ensure real-time input of sensor data.
• Configure asynchronous pipelineEnables the Foresight Reasoning asynchronous reasoning mechanism, allowing the model to predict the next state in parallel while executing the current action segment.
• Closed-loop calibrationSet up a real observation feedback channel so that the model can recalibrate its predictions each time a new observation is returned, thus avoiding physical drift.
• Task adaptationIf a new task needs to be performed, a small amount of robot operation data is collected, and the data is then finely tuned based on the pre-trained weights.

The core advantages of LingBot-VA 2.0

• IncarnationFrom architecture and data to training objectives, everything is tailored to the physical world, rather than being a fine-tuning of a digital video generation model.
• Real-time and efficientSingle GPU 150Hz inference meets the requirements of real-time closed-loop control.
• Cause and effect consistencyStrict unidirectional temporal modeling avoids the loss of motion accuracy and catastrophic forgetting caused by bidirectional attention.
• Data efficiencyMCP multi-step target prediction improves training convergence speed by 2.3 times and reduces reliance on robot labeled data.
• Full-stack collaborationIt forms a complete closed loop of perception-prediction-execution with LingBot-Depth, LingBot-VLA, LingBot-Video, etc.

The project address for LingBot-VA 2.0

• Project official websitehttps://technology.robbyant.com/lingbot-va-v2
• Technical Papers: https://github.com/Robbyant/lingbot-va/blob/main/LingBot_VA2_paper.pdf

Comparison of LingBot-VA 2.0 with similar competing products

Application Scenarios of LingBot-VA 2.0

• Home service robotsPerform long-term household chores such as tidying up the desktop, putting things back in their place, and setting out tableware. Utilize long-term memory and arm coordination to complete complex daily tasks.
• Industrial Dynamics CaptureThis technology enables real-time grasping of moving targets on conveyor belts and assembly lines. The model predicts the future position of the object and synchronizes the movement rhythm, replacing the traditional photoelectric triggering solution.
• Precision assembly operationsIt enables flexible force-controlled gripping and assembly of chips, wafers, and fragile components, relying on fine-grained visual servoing to protect the integrity of materials.
• Human-computer interaction entertainmentIt supports high-frequency interactive scenarios that require real-time prediction of opponent's movements, such as ice hockey matches and tabletop games, enabling millisecond-level reaction and strategy adjustment.
• Warehouse logistics sortingIt can sort, handle and palletize goods in a dynamic and changing warehousing environment, and adapt to the generalized grasping needs of packages of different sizes and shapes.