Chinese Robot Firms Adopt External AI Models While Retaining Core R&D Efforts

Report Summary

How much less R&D would robot companies have to do if they didn't train base models from scratch? LeJu borrows Ant Lingbo's model to train industrial operations, and Xinghaitu is simultaneously developing and using external models. Technical achievements are flowing across enterprises, but for robots to truly do work, they still need to train themselves.

On November 5, robot models developed by three Chinese companies appeared in the same open-source software update list.

They are respectively Xinghaitu's G0.5, Ant Lingbo's LingBot-VLA 2.0, and Ziyuanbian's Wall-OSS-0.5; although developed by different teams and with different technical architectures, they have now all received support from the new version of LeRobot 0.7.

This looks like just a software update message. But before this list appeared, some more interesting things had already happened between Chinese robot enterprises.

In August this year, Leju released a robot model for industrial operations, KUAVO-VLA-1.0. In a real-machine evaluation titled “Schaeffler Small Parts Loading,” its task success rate reached 60%, while a control scheme using the same base model and receiving the same targeted task training had a success rate of only 27%.

The interesting part is that these two sets of solutions actually have the same starting point.

Leju did not start training its own base model from scratch, but instead adopted Ant Lingbo's LingBot-VLA 2.0, and then continued training it with its own 680 hours of industrial operation data, finally obtaining a model more suitable for its own robots.

Another company developing robot models, Xinghaitu, has an even more complex situation. It not only develops its own models but also provided hardware, data, and testing support to Ant Lingbo, and is now integrating models from other companies into its own development platform.

Robot enterprises that originally appeared to be competing individually have begun to use each other's results in different R&D stages.

This brings up a question worth asking: when robot companies can already use “brains” developed by others, what do they still need to research and develop themselves?

How can a set of external models become Leju's own robot capabilities?

Leju did not start contacting external models only in August of this year.

On May 30, its open-source project LeTools-Learning announced support for multiple existing robot models, including the π series, GR00T, and LingBot-VLA.

For a company manufacturing humanoid robots, being able to connect to more models means developers have the opportunity to try different algorithms on their own robot hardware. However, these models are not directly usable after being downloaded.

Kuavo has its own joints, cameras, end effectors, and control systems. Different models use different data structures, action representations, and reasoning programs.

For example, a piece of dual-arm operation data, in addition to recording what the robot saw, also needs to include the positions and action changes of the left and right arms, as well as the correspondence between this information and the camera images. If a different robot is used, where the joint configuration or end effector is different, the original data processing method may need to be adjusted.

LeJu therefore specifically developed its own toolchain.

In the public LeTools-Learning project, different modules such as data conversion, model training, inference services, and robot deployment can be seen. Developers need to organize the raw ROS data collected by Kuavo into datasets that the model can read, and then complete the corresponding adaptation based on the specific model, single-arm or dual-arm configuration, and end effector.

These tasks did not disappear because of the use of external models.

However, Leju also did not choose to train a set of general models from scratch every time. The KUAVO-VLA-1.0, launched on August 26, directly uses Ant Lingbo's LingBot-VLA 2.0 as the base.

Next, Leju did something closer to the robot's own business: using its own real machine data, it let the external model further learn industrial operations.

It accumulated 680 hours of dual-arm teleoperation data, covering 100 industrial tasks, including part grasping and releasing, sorting, assembly, packaging, and tool operation, forming a total of about 58,800 operation demonstrations.

These data are not factory videos filmed by staff holding cameras, but rather humans using remote-controlled equipment to control real Kuavo robots to complete tasks. During data collection, the system simultaneously records the images from the robot's head and wrist cameras, as well as information such as the robotic arm status and operation trajectories.

Subsequently, the collected data undergoes manual review, format conversion, and automatic filtering before entering the training system.

For the model, these data provide not only the appearance of objects but also the experience of how a specific robot completes an operation.

Leju first uses the data from these 100 tasks to perform industrial domain retraining on LingBot-VLA 2.0, attempting to let the model acquire industrial operation capabilities more suitable for Kuavo.

But it does not end here. When a robot truly faces a specific job, Leju will still conduct post-training around that task.

This is equivalent to dividing the R&D work into different levels: the base model provides the already learned visual, linguistic, and motor capabilities, industrial domain training makes it familiar with a category of work, and the final specialized training allows the robot to further adapt to specific tasks.

Leju did not remanufacture all capabilities, nor did it treat external models directly as finished products; it chose to continue building its own industrial operation capabilities on the existing foundation.

The question is: just how much change does this approach actually bring?

The same “brain,” some tasks have been learned, some still have not

Leju released a real-machine evaluation scorecard of 25 operational tasks.

Among them, 20 belong to tasks covered by previous industrial field training data, while the other 5 come from the public GM100 task set and did not enter the first stage of training data.

Therefore, this is not a zero-shot test of letting a robot receive a strange task and trying it directly, but rather a comparison of the performance of different models on specific tasks after corresponding training.

“Schwerdtauer small parts loading” is one of them. According to the task description in the Leju technical report, the robot needs to grab parts from a bin and then place them on a conveyor belt.

The action does not sound complicated, but the robot needs to complete a series of operations from identifying the target and controlling the grasp to moving and placing it; failure in any one link could make the entire task impossible to complete.

According to the results released by Leju, KUAVO-VLA-1.0, based on LingBot-VLA 2.0, reached a complete task success rate of 60%. As a comparative solution, specialized training directly from LingBot-VLA 2.0 had a success rate of 27%.

In other words, under the same set of evaluation conditions, the model that underwent retraining in the industrial field showed a significantly higher proportion of completing this task.

Looking further, the process scores for the two models were 88% and 52% respectively. The process score statistics count how many steps the robot completed during the operation; even if the final task was not fully completed, the parts already completed will be counted.

This set of results is very representative.

Industrial operations are often not a simple single grasp. A robot may need to identify a target, grab a part, adjust its position, and then place it in a designated area. Failure in any one link could make the complete task impossible to finish.

However, Leji did not release a sufficiently detailed step-by-step failure analysis, so it is not yet possible to determine exactly on which action the model achieved improvement.

Moreover, if one continues to look through this report card, some not-so-pretty results will also be found.

In the “Part Size Sorting” task, the complete task success rate of KUAVO-VLA-1.0 remains at 0%.

In the “fast-moving consumer goods manual loading” task, it achieved a success rate of 60%, which is lower than the 80% of another model participating in the control, π0.5.

In other words, models retrained in the industrial field do not perform best on all tasks.

Even the same set of models may exhibit significant capability gaps between different industrial operations.

Therefore, for Leju, borrowing Ant Lingbo's foundation model is only part of the R&D work; a more practical question still needs to be answered: how to make this set of models learn to use Kuavo's hands to complete the industrial tasks the company hopes to enter?

680 hours of data, industrial field retraining, post-training for specific tasks, and the robot's own control adaptation are all centered around this question.

Moreover, these tasks may not necessarily be finished in a single completion.

The success rate of one task is improved, while another task may still fail. Even for similar pick-and-place operations, as long as the parts, operation sequences, or working conditions change, new training and validation may be required.

Leju's experiments also cannot prove exactly how much R&D cost was saved by adopting external models, or whether the retrained models can already stably enter factories.

But it at least provides a real development case: robot companies do not have to develop all model capabilities from scratch, yet they still need to use their own data and engineering work to turn external models into capabilities suitable for specific robots.

In the past, people often understood the technological competition of robotics companies as who possesses a stronger “brain”.

Now, at least in Leju's practice, where the foundation model comes from and what work the robot can ultimately complete can already be two different questions.

However, Leju is still primarily a user of external models.

In another company, the relationship between the model developer and the user is no longer so simple.

Who is the model developer, and who is the user?

In February this year, when Ant Lingbo open-sourced the early LingBot-VLA model, Xinghaitu appeared in its technical cooperation list.

This was not a simple software adaptation. According to public reports, Xinghaitu provided technical support for LingBot-VLA in aspects such as data support, evaluation and verification, and cross-robot body deployment. Its R1 Pro robot was also one of the important hardware platforms used for the research and verification of this model.

In other words, Ant Lingbo's development of a foundation model oriented toward different robots required the help of Xinghaitu's real robots and relevant technical conditions for verification.

This point is not difficult to understand. Robot models differ from language models that only process text. The actions it outputs must ultimately be completed by robotic arms, joints, and end effectors. Whether the model can work on robots with different configurations requires real hardware to provide verification conditions.

Even if a company has strong model R&D capabilities, it does not necessarily need to manufacture all the robots used for experiments itself.

And the special point of Xinghaitu is that it does not just provide hardware for others.

In June this year, Xinghaitu released its self-developed embodied intelligence base model G0.5. At the same time, it is also building its own robot development platform EDP.

In the public introduction of the EDP platform, in addition to self-developed models such as G0.5, it also listed external models such as Physical Intelligence's π0, π0.5, as well as Ant's LingBot-VLA.

Xinghaitu hopes that developers can complete tasks such as data management, model post-training, deployment, and evaluation on the same platform. According to information obtained, Xinghaitu has already included these models in the scope of support, but has not further disclosed whether each model has undergone complete third-party real-machine verification or has already formed large-scale commercial use.

But even so, this set of facts is already interesting enough.

When Ant Group's Lingbo is developing models, Xinghaitu can provide robot hardware, data, and evaluation support.

In developing its own G0.5, it becomes the provider of the base model.

And in the EDP platform, it also includes models developed by other companies into the supported scope.

If you look back at LeTools-Learning from Leju, you will find that Xuhaitu's G0.5 is now also listed in its external model support list.

This forms a rather special relationship.

The Ant Lingbo model can become the basis for the Leju Industrial model.

Xinghaitu's robot hardware, data, and testing conditions can participate in the development and verification of Ant's Lingbo model.

Xinghaitu's self-developed models also enter LeJu's toolchain and LeRobot's public development ecosystem.

Enterprises that might originally have been simply categorized as model companies, robot manufacturers, and tool platforms are actually exchanging roles in different R&D stages.

Of course, this does not mean these companies have already jointly built a unified model system.

But these relationships reveal at least one thing: in the robotics industry, developing one's own models and using others' models are not contradictory; providing robot hardware and participating in model R&D are also not necessarily two unrelated businesses.

A company can obtain technology from others and also become a technology provider that others' R&D relies on in another link.

This is clearly different from the simple division of labor people are familiar with, where “model companies develop brains and robot companies purchase brains.”

So, when different companies are all developing models, using models, and also building their own tools, does the industry still need a set of common development infrastructure?

The LeRobot update on November 5th happens to provide an observation window.

A common toolbox cannot still hold all the work

LeRobot 0.7 is not the starting point for robotics companies to begin using external models.

Leju had already publicly supported multiple external models as early as May and released its own industrial vertical domain model in August; the cooperation between Xinghaitu and Ant Lingbo occurred even earlier at the beginning of the year.

But these originally scattered technical relationships are having more intersections in public tools.

In this update, there is one change that particularly illustrates what problems the jointly developed tools are solving: robot training data no longer needs to be fully downloaded to the local machine beforehand.

In the past, a large-scale set of robot manipulation data often contained a large amount of camera videos and corresponding action records. To use this data to train models, developers might first need to solve the problems of how to store, download, and read them.

LeRobot 0.7's new streaming reading mechanism can read the actual part of video content needed from a remote dataset based on training progress, without having to move the entire dataset to the local machine first.

For different development teams using the same data format, this means they can reuse existing data reading and training organization tools, without having to develop a set of programs for processing large video data from scratch.

This capability cannot guarantee that every project will achieve the same training acceleration, nor can it solve data collection and quality issues for enterprises. But it does indeed hand over a portion of the repetitive software engineering work to commonly maintained tools.

The same set of tools can now also support more models from different companies.

This means that more and more development work can start from models and software already completed by others.

But the practices of Leju and Xinghaitu also illustrate the boundaries of this reuse.

A set of public tools can help developers organize training, but it will not automatically know how the Kuavo dual-arm robot should grasp a part.

An external foundation model can provide existing operational capabilities, but it cannot guarantee that another robot will achieve the same results directly in a new task.

Even if models, data, and tools can be shared, the embodiment control, task training, and test verification specifically required by the robot still need to be completed.

Leju uses Ant Lingbo's models and invests its own industrial data and training systems; Xinghaitu develops G0.5 and also provides hardware, data, and evaluation support for other companies' models.

What is worth paying attention to now is not just how many companies have open-sourced robot models, but what capabilities can actually be developed after these models enter other companies' R&D systems.

Leju's report card of 25 tasks provides an imperfect, yet very specific answer.

The same base model, after being trained in the industrial field and specifically developed, achieved better results on some tasks; but for other seemingly similar industrial operations, it still cannot complete them fully.

Robot companies can borrow others' already trained “brains,” but they cannot use this to skip the process of letting the robot truly learn to work.

Where models come from is becoming a choice; what robots can ultimately do still depends on individual enterprises continuing to verify them on real hardware and specific tasks.

END