LingBot-VLA 2.0 - A new generation of embodied intelligent base model open sourced by AntLingbo.

What is LingBot-VLA 2.0?

LingBot-VLA 2.0 is a new generation of embodied intelligent base model open-sourced by Antminer Technology. Based on 60,000 hours of pre-training data, including 50,000 hours of real-machine data and 10,000 hours of human operation data, it covers 20 robot configurations from 17 brands and supports single-arm/dual-arm, bipedal/wheeled, and multi-degree-of-freedom movements such as head, waist, and chassis. The model employs a sparse MoE architecture and a unified motion space, outperforming π0.5 and GR00T N1.7 in both GM-100 dual-arm evaluations and long-range mobile operation tasks, with an RTX 4090 inference time of less than 130 milliseconds.

Main functions of LingBot-VLA 2.0

• 20+ configuration generalizationSupports 20 robot configurations from 17 brands, covering single-arm, dual-arm, bipedal, and wheeled forms.
• Multi-degree-of-freedom controlIt supports full-body freedom, including the head, waist, end effector, chassis, arm, and gripper.
• Dual-arm coordinated operationIn the GM-100 evaluation, it leads the competition in task progress score and success rate.
• Long-range mobile operationsSupports the advancement of long-sequence tasks across different domains, such as refrigerator organization and stove cleaning.
• Efficient inference deploymentThe post-trained version kept inference time under 130 milliseconds on an RTX 4090.

Technical Principles of LingBot-VLA 2.0

• Large-scale heterogeneous data pre-training60,000 hours of high-quality data were cleaned from 90,000 hours of raw data, including 50,000 hours of real machine operation from 20 robot configurations and 10,000 hours of first-person human operation videos, covering a variety of physical interaction scenarios.
• Unified Action Space RepresentationAlign 20 heterogeneous robot configurations to shared action dimensions such as Arm, EEF, Gripper, Move, Waist, Head, and Hand to achieve scalable training across ontology.
• Sparse MoE architectureThe model adopts a sparse hybrid expert architecture with 1.6B total parameters and 0.6B activation parameters. Under the same activation parameter budget, the training loss and validation action error are significantly lower than those of the Dense 0.6B model, balancing performance and inference efficiency.
• Dual-query distillation perception predictionThe Dual-query Distillation mechanism simultaneously predicts current and future RGB, depth maps, and DINO visual representations, enhancing the model's ability to understand the temporal sequence of dynamic environments.
• Multi-stage data cleaning pipelineThe raw data is divided into robot stream and egocentric stream, and the training data is gradually refined through stages such as smoothing motion and state filtering, high-quality video screening, VLM filtering, hand trajectory reconstruction and standardization.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use LingBot-VLA 2.0

• Obtain model weightsVisit Hugging Face or Moda Community to download the LingBot-VLA 2.0 pre-trained model.
• Cloning open source codeGet the inference and deployment code from GitHub, and read the README to understand the environment dependencies.
• Configure the robot bodyPrepare the corresponding robot hardware and sensors based on the official list of 20 supported configurations.
• Deployment reasoningLoad the model onto the RTX 4090 or an adapted edge computing platform such as Sunrise S600, Jetson Thor/Orin.
• Execute operation tasks: Send commands to the robot through a unified motion interface to complete dual-arm operations or long-distance movement tasks.

The core advantages of LingBot-VLA 2.0

• The strongest cross-configuration generalization abilityIt covers 20 robot configurations and is one of the open-source VLA models in the industry that supports the most types of bodies.
• Dual-arm and mobile operation leadingIn both the GM-100 dual-arm evaluation and the long-range mobile operation cross-domain evaluation, it surpassed π0.5 and GR00T N1.7.
• Inference efficiency optimizationThe sparse MoE architecture enables only 0.6B parameters to be activated in a 1.6B model, achieving inference times of less than 130ms on an RTX 4090, making it suitable for edge chips.
• Ecosystem adaptation is perfectWe have completed the adaptation of mainstream chips such as Digua Sunrise S600 and NVIDIA Jetson Thor/Orin, and have established cooperation with chip manufacturers such as Leju and Titanium Tiger.

The project address for LingBot-VLA 2.0

• Project official websitehttps://technology.robbyant.com/lingbot-vla-v2
• GitHub repositoryhttps://github.com/robbyant/lingbot-vla-v2
• HuggingFace model libraryhttps://huggingface.co/collections/robbyant/lingbot-vla-v2
• arXiv technical paper: https://arxiv.org/pdf/2607.06403

Comparison of LingBot-VLA 2.0 with similar competing products

Application Scenarios of LingBot-VLA 2.0

• Retail servicesIt can perform dual-arm operations such as shelf organization, product retrieval and inventory counting in pharmacies, supermarkets and other settings.
• Logistics warehousingIt performs processes such as parcel sorting, cargo handling, and long-distance mobile loading and unloading, and is compatible with wheeled mobile chassis.
• Industrial manufacturingIt is used to complete tasks that require high precision and multi-degree-of-freedom coordination, such as assembling precision parts, picking and placing tools, and inspecting equipment.
• Home servicesSupports everyday household chores such as refrigerator organization, stove cleaning, and joint object operation.
• Scientific research and developmentIt provides an open-source VLA platform for universities and research institutions to accelerate the iteration of embodied intelligence algorithms and ontology adaptation research.