LDB Embodied Intelligence General-Purpose Robot Released, Driving Industrial Scene Implementation

A single product is placed in a different orientation, and pickers usually only need to adjust the angle of their reach. However, robots may need to re-evaluate how to approach, from where to grasp, and what will happen after picking it up.

These small changes that repeatedly appear in warehousing constitute a question that is often overlooked when embodied intelligence is applied to industrial scenarios: The tasks may seem the same, but the execution conditions may not be the same.

On October 22, the 2026 Kuaishou Global 1024 Developer Festival was held at the Beijing National Convention Center. Kuaishou's incubated embodied intelligence company Lingdong General brought the LDB Embodied Intelligence General-purpose Robot, simulating a small-item sorting scenario in a warehouse.

It targets the most mundane type of work in B2B warehousing and logistics—From daily necessities, 3C electronics to shoes, clothing, cosmetics, items of various shapes are squeezed in storage locations in various postures. The robot must navigate through shelves and aisles, identify the target, and complete the picking and placing.

Factories are not unfamiliar with robots. According to the International Federation of Robotics (IFR) September release of the World Robotics 2026 report, approximately 603,000 new industrial robots were installed in global factories in 2025, representing an 11% increase year-on-year. Among these, approximately 354,000 were in China, nearly 60%. Of the robots installed in China, domestic brands accounted for 55%.

Traditional automation excels at repeating tasks in processes with clear rules, but when product categories, placement positions, or workstation layouts frequently change, reprogramming and on-site modifications often increase adaptation costs. Embodied intelligence aims to further take over these tasks that require flexible adaptation.

These tasks are mostly labor-intensive and increasingly difficult to recruit workers for. Ji Chao, founder and CEO of Lingdong General, previously envisioned a production line ten years from now: robots becoming execution units, working in high-risk, dirty, noisy environments, taking on heavy physical labor; frontline workers would shift to planning, production scheduling, and human-robot collaboration. This human-robot division of labor is a direct illustration of how domestic AI creates social value.

But trust issues also arise. Embodied robots have greater autonomy in decision-making and can adjust their actions according to the situation on site. However, industrial customers must also confirm whether the robot can reliably complete tasks in complex environments and whether it can be effectively constrained when its actions may pose risks. The more autonomous a robot becomes, the higher the requirements for its reliability in industrial settings.

To help industrial customers cross the 'dare to use' threshold and accelerate the practical application of embodied intelligence, during this Developer Festival, Lingdong General released the iFlyBot-WAM industrial world action model and the LindenBot-Harness execution organization system, and also unveiled the LDU environment-based data collection solution.

From consequence prediction to task execution, and then to continuous iteration driven by real data, Lingdong General is leveraging the source AI technology of iFlytek to respond to a real-world question through full-stack self-research: When embodied robots start to make their own judgments and actions, what makes industrial customers dare to entrust tasks to them?

Source AI technology as the foundation: first predict consequences, then decide to proceed

In industrial settings, flexibility often means an additional cost.

Traditional robotic arms can achieve high efficiency and stability by repeating tasks on fixed workstations. However, in multi-product warehouse scenarios, the size, position, and orientation of items frequently change. If each situation requires reprogramming and repeated debugging, the cost of automation upgrades will rise accordingly.

Embodied large models give robots more flexible adaptability, but also bring new challenges. For example, a facial cleanser leaning against a shelf, even if the robot accurately identifies the target, it still needs to determine where to start, how to grasp it, and whether it will knock over adjacent items.

When robots start to decide how to perform tasks, the uncertainty behind actions becomes a risk that industrial customers must consider.

Ji Chao previously mentioned in an interview that There is still a significant gap between embodied intelligence and industrial-level applications, with one important issue being the model's misjudgment in unfamiliar scenarios. In a factory, a single incorrect operation may cause damage to goods, or even disrupt normal operations. How to enable robots to flexibly respond to changes while avoiding unreasonable actions has become a direction that Lingdong General continues to focus on.

Last year's 1024 Developer Festival, Lingdong General and iFlytek jointly released the iFlyBot-VLM visual language model and the iFlyBot-VLA visual language action model: the former enables robots to understand their surroundings, while the latter translates task instructions into specific actions. Relying on iFlytek's accumulated source AI technologies in large models and multimodal perception, Lingdong General continues to advance the self-research of embodied models, building perception, understanding, and action generation capabilities tailored for industrial scenarios.

The recently released iFlyBot-WAM industrial world action model further enhances the predictive capabilities of the physical world on this foundation. According to Lingdong General's technical design, VLA generates actions based on visual information and task instructions, while the world model predicts the possible state of objects and environmental changes resulting from these actions, such as whether surrounding items will be affected after picking up this product.

However, prediction alone is not enough to fully reassure industrial customers; it also requires judging whether the action is suitable for execution.

For this purpose, iFlyBot-WAM introduces the 'World Model Evaluation + Deterministic Safety Arbitration' mechanism. The world model first evaluates the potential outcomes of an action, and then explicit safety rules determine whether to approve, adjust, or reject the action. The model is responsible for prediction, while the rules decide approval—this division of responsibilities provides robots with a clearer safety boundary for autonomous operations, allowing industrial customers to more reliably assess operational risks.

For industrial customers, another layer of value of iFlyBot-WAM lies in deployment efficiency. It emphasizes the ability to generalize across perspectives and layouts, meaning that when the location of goods or shelf layout changes, the robot is expected to reuse previously learned operational capabilities, reducing the need for reprogramming and on-site debugging. The time customers need to get the robot to adapt to a new environment will directly impact the application cost of embodied intelligence.

But in real operations, a single action's safety and reliability is just the beginning. From identifying goods to moving, grasping, and placing them, the robot still needs to complete a series of consecutive tasks. How the judgments formed by the model can be further converted into stable execution capabilities still tests the overall coordination level of the robot system.

Creating social value with domestic AI: Starting from stabilizing an entire shift

Industrial customers will not purchase robots solely for a single successful demonstration.

In warehouse picking, any issue in the process from identifying goods to grasping, moving, and placing them may affect the completion of the entire task. Customers are more concerned about whether a robot can maintain a stable work rhythm throughout a shift, how much manual intervention is needed if a fault occurs, and whether it is cost-effective in the long term.

Ji Chao has repeatedly mentioned that industrial scenarios follow a “cost-benefit logic,” and what customers care about is actually quite simple: “Can it solve the problem, how long will it take to break even, and can it run stably?” In July this year, he once revealed that Lingdong General has already compressed the break-even cycle to within two years in some scenarios, while industrial customers typically accept a break-even cycle of two to three years.

This account needs to be settled, and for robots to have the capability to continuously complete tasks. Large models can understand instructions and plan actions, but in actual operations, they still need to coordinate different operational steps and constantly judge whether the task is completed according to the plan.

The newly released LindenBot-Harness execution organization system is precisely to solve this problem. Among them, the agent (Agent) is responsible for task decomposition and multi-objective coordination planning; Harness organizes the execution plan according to the on-site context, uniformly schedules perception, operation, navigation, and motion capabilities, and checks the execution results, embedding safety control into the operational process.

Previously, such operational processes largely relied on engineers pre-writing programs and rules. When switching to a new warehouse, facing different product layouts and operational requirements, it often required re-calibration.

Harness aims to allow robots to autonomously adjust the execution process according to on-site conditions, reducing reliance on item-by-item programming. For enterprises that need to deploy robots in different scenarios, less repeated development means lower delivery costs.

However, no matter how well the execution process is arranged, it still needs robots that can withstand actual working conditions. Previously, in logistics sorting sites, Lindong General encountered issues where model inference could not keep up with the production line pace, and the end-effector solution combining suction cups and grippers also struggled to simultaneously adapt to light and soft packages and heavier packages. The team then spent half a year at the customer site, making coordinated adjustments to the model architecture, data system, and end-effector gripping solution.

This reveals an overlooked complexity in industrial settings: Model inference needs to keep up with the operational pace, action planning must consider the execution conditions of the hardware, and mechanical structures also need to adapt to the operational requirements of different items. Any weakness in any of these links could affect the final operational effectiveness.

Today's LDB has been gradually refined through such repeated verification. According to Zhang Jiyang, the joint founder and product VP of Lindong General, the team began defining product requirements in 2024, iterated four generations of laboratory functional prototypes, and the one exhibited in August this year is the fifth generation. The entire machine was designed from the outset for industrial scenarios, with targeted design for complex conditions such as high temperature, high humidity, and electromagnetic interference.

On May 27th this year, LDB received the CR certification from the National Robot Inspection and Evaluation Center (Headquarters), making Lingdong General the first company in the country to obtain this certification in the industrial application field. During a 9-month research and testing period, the product passed multiple tests including electromagnetic compatibility, electrical safety, and mechanical safety, and also obtained the EU CE-MD access qualification.

These certifications provide verifiable safety and compliance basis for robots entering the industrial market, and also reflect Lindong General's integration of industrial standards into product design from the research and development stage.

For customers, whether a robot can run stably for a long time ultimately depends on its performance in real working conditions, where it must be tested for operational efficiency, failure rate, and maintenance costs. Only when a robot can stably complete an entire shift can it potentially take over those high-intensity, hard-to-recruit positions, and the envisioned human-machine division of labor by Ji Chao also has a foundation.

From Harness organizing tasks to LDB itself completing operations, Lindong General is transforming the model's intelligent decision-making into sustainable industrial capabilities. And when robots enter more factories, facing different products and operational environments, how to reduce repeated debugging and let existing experience support the next deployment, poses new requirements for data accumulation and model iteration.

Accelerating the implementation of embodied intelligence: enabling experience to be reused across more factories

A solution that works in one factory may encounter different products, workstations, and operational processes in another. If each new project requires engineers to stay on-site for a long time, re-collecting data and repeatedly debugging, the number of deliveries increases, but the cost may not necessarily decrease.

Whether existing experience can be reused determines the efficiency of scaling up the delivery of embodied intelligence.

Lindong General attempts to solve this issue from the data side. The LDT data collection, training, and inference all-in-one machine, unveiled in June 2025, has already integrated the data collection, model training, and inference process, and achieved data interoperability with LDB. According to the company's previous disclosure, the team has conducted nearly thirty factory surveys and carried out real machine deployment and regular data collection in over ten leading logistics and automotive companies.

However, industrial data is often held by customers and is difficult to obtain through public channels. Traditional data collection methods relying on the robot itself face issues such as high equipment costs and limited deployment.

The newly unveiled LDU is Lindong General's new attempt to further reduce the cost of data acquisition. It adopts a replaceable battery backpack design, supports first-person and dual-hand data collection, and simultaneously records multi-modal information such as vision and motion trajectories. The data collection process does not require the robot itself, enabling it to cover more real-world scenarios at a lower cost.

How can on-site feedback be transformed into effective data? Lingdong General has already made more specific attempts. At the World Manufacturing Conference in September this year, Lingdong General demonstrated LDB's 'human-machine-in-the-loop' capability: when a robot encounters a situation it cannot handle autonomously, the operator can take over using a VR device, and the system simultaneously records motion and visual information.

These data can be used to train and optimize models, which after validation and re-deployment help robots handle situations they previously struggled with, thereby reducing the reliance on human intervention for similar tasks.

By this point, Lingdong General has gradually formed a collaborative closed loop of “data + model + execution + ontology” four-layer capabilities.

In this loop, LDT expands data supply along with future LDU, iFlyBot-WAM enhances model capabilities through training, Harness is responsible for organizing tasks, and the LDB at the ontological layer completes operations on-site and provides feedback. Issues identified during execution can serve as a basis for model optimization in the next round, and also support RSI's continuous learning and self-evolution closed-loop.

The value of full-stack self-research lies in the ability of each layer to iteratively improve around the same on-site problem. Model training demands can reverse-guide data collection, and feedback from actual operations can drive simultaneous adjustments to model strategies, execution processes, and ontology design. The accumulated experience in collaborative development and verification over the long term forms the core technical barrier of Lingdong General.

The full-stack self-controlled starfire large model and related industrial resources from iFlytek provide technical and ecosystem support for Lingdong General, enabling it to focus its R&D investment on key areas that directly impact industrial operation effectiveness.

On production lines where it's increasingly difficult to recruit personnel, creating social value with domestic AI ultimately comes down to a series of robots that can reliably perform tasks. This is also the starting point for Lingdong General's accelerated push for embodied intelligence deployment. Whoever can make the experience accumulated in one factory become the capability of the next factory will have a greater chance of seizing the initiative in industrial embodied intelligence.