Ant Group's Lingbo Signs Multiple Enterprise Deals, Embodied AI Brain Accelerates onto Production Lines

Embodied intelligence is moving from the lab to real production lines. Ant Group's embodied intelligence company Lingbo has been making frequent moves recently, successively entering into deep strategic partnerships with multiple smart manufacturing and modern logistics enterprises including Longsheng Weirui, Dema Technology, and Deer Xiangwang, exploring deployment around scenarios such as industrial loading/unloading, automotive manufacturing digitalization, and flexible logistics sorting.

1. Signing Trends: Industry and Logistics Advancing on Two Fronts

According to public information, Lingbo's partners this time cover two major sectors: industrial manufacturing and smart logistics. Longsheng Weirui, a subsidiary of Longsheng Technology, will jointly build an "Embodied Intelligence Robot Innovation Center" with Lingbo, initially focusing on industrial loading/unloading scenarios, with future clients including global industrial manufacturing leaders such as Bosch and Baosteel. Deer Xiangwang will explore digital transformation pathways for automotive manufacturing workshops together with Lingbo.

On the logistics side, Dema Technology, as a logistics equipment and solutions provider, will collaborate with Lingbo on engineering enablement, secondary development coordination, and new technology testing and validation, with the core goal of solving the challenge of high-speed flexible sorting of massive parcels. These partnerships are not merely framework-level agreements but specify concrete scenarios and delivery directions, demonstrating Lingbo's determination to advance industrial deployment.

2. One Brain, Multiple Robots: Solving the Adaptation Cost Challenge

Embodied intelligence's path to industrial application has long been constrained by adaptation costs. When a robot switches from one scenario to another, it often requires re-collecting data, retraining models, and debugging systems; replacing hardware may also require entirely new adaptation work. Insufficient model generalization combined with hardware differences makes it difficult to replicate single-point validation across more production lines.

Lingbo CEO Zhu Xing has pointed out that the company broke away from the traditional approach of single-machine tuning, choosing instead to train an embodied foundation model for the physical world from scratch, building a native general-purpose "brain" for robots. Its self-developed LingBot-VLA 2.0 foundation model has already been adapted to 20 robot configurations from 17 brands during the pre-training phase, establishing the industry's first "one brain, multiple robots" engineering benchmark. The core value of this strategy lies in: the same brain can drive different configuration bodies, reducing migration costs across scenarios and hardware.

3. At the Industry Expo: Real Multi-Scenario Operations Demonstrated

At the ongoing China International Industry Fair, Lingbo showcased real-world applications including precision pharmacy picking, smart logistics flexible sorting, and industrial production line loading/unloading alongside multiple ecosystem partners. Two robot bodies with different configurations from Tihu and Stardust Intelligence were visible on-site, both equipped with the Lingbo Brain LingBot-VLA 2.0, completing high-precision industrial loading/unloading and flexible sorting tasks respectively.

This "different bodies, same brain" demonstration directly validates the generalization capability of one brain, multiple robots in real industrial environments. Previously, Lingbo had already proven a continuous pharmacy picking solution in offline pharmacies, accumulating engineering experience from narrow-aisle navigation to continuous operations.

4. Open-Source Ecosystem and Global Layout

Lingbo's open-source strategy provides another growth curve for its technological influence. As of public data, its open-source models have accumulated 32,900 stars on GitHub, ranking among the top domestic embodied models. Major overseas technical communities have completed adaptation of Lingbo's models: LingBot-VLA and LingBot-VA have both completed ecosystem adaptation for LeRobot, the world's largest robot development platform; LingBot-Vision has been merged into the main branch of the mainstream vision backbone model library timm; and LingBot-World has become the world's first world model to simultaneously receive support from both vLLM and SGLang, the two mainstream open-source inference engines.

On the globalization front, Lingbo has reached cooperation with the Arab Digital Economy Alliance in areas including technical models, localized scenario adaptation, and industrial ecosystem integration, opening a new phase for the large-scale deployment of "one brain, multiple robots" in the global physical world.

LeRobot is an open-source robot learning framework launched by Hugging Face, providing a unified data format, pre-trained model library, and training toolchain, regarded as the "model and dataset hub" in the robotics field; timm (PyTorch Image Models) is the most widely used backbone network library in computer vision, containing numerous mainstream vision model implementations; vLLM and SGLang are currently the two mainstream open-source engines on the large model inference side, with the former known for high-throughput PagedAttention scheduling and the latter excelling in structured generation and multi-turn reasoning efficiency. Lingbo's models have respectively completed adaptation and integration with these ecosystems, meaning developers can directly call its capabilities within mainstream toolchains without building environments from scratch, which is crucial for its technology diffusion and ecosystem influence.

5. Industry Observation: Embodied Intelligence Enters the Engineering Delivery Phase

From narrow-aisle navigation in smart pharmacies to production line operations at the Industry Expo, to densely signing benchmark orders in industrial manufacturing and smart logistics, Lingbo is advancing commercialization through a dual-wheel drive of "leading models + engineering tools." The focus of competition in embodied intelligence has shifted from model parameter comparisons to stability and generalization capability in real scenarios.

Whether it can continuously deliver in demanding industrial and logistics environments will determine the final landscape of this wave of embodied intelligence deployment. Lingbo's successive signings provide an observation sample for the industry: a general-purpose brain plus ecosystem collaboration may be a viable path to large-scale adoption.

The so-called "engineering delivery phase" refers to the stage relative to embodied intelligence's previous focus on paper metrics and demonstration videos. Early industry competition centered on laboratory metrics such as model parameter scale and benchmark success rates, but real production lines require stability under long-term continuous operations, controllable failure rates, and the ability to interface with existing PLC, MES, and other industrial systems. These metrics are difficult to solve through single-point technological breakthroughs; they require scenario data accumulation, hardware adaptation, and delivery team coordination. Therefore, the decisive factor at this stage has shifted from "how strong the model is" to "whether stable delivery is possible," which also explains why Lingbo is densely signing benchmark industrial and logistics clients—orders themselves serve as validation channels for engineering capability.

Background Supplement

VLA (Vision-Language-Action) models are a class of foundation models in embodied intelligence whose core idea is to unify visual perception, language instructions, and action generation into a single model: the model receives camera images and natural language instructions and directly outputs action sequences for robot joints or end effectors. Compared to traditional separated "perception—planning—control" pipelines, the VLA end-to-end architecture enables robots to better understand semantic information in open scenarios and possesses cross-task generalization potential. LingBot-VLA 2.0 belongs to this technical approach, and its "training from scratch for the physical world" means that training data comes largely from real robot operation trajectories rather than pure internet data, which is also the prerequisite for its ability to adapt to multiple robot configurations.