Ant Nest Lingbo: Why train a robot’s brain from scratch?

Throughout the entire briefing session, the media kept asking the same question in many different ways.

How far is today’s robot “brain” from truly entering the physical world to actually work?

The assessment given by Ant Lingbo CEO Zhu Xing and Chief Scientist Shen Yujun is far more level-headed than the market’s heat. Today’s robot “brain” may not have reached the GPT-1 moment yet. The industry has not yet seen a real wave of intelligence emerge, and the technical roadmap is still far from converging.

Over the past year, concepts such as VLA, world models, and video action models have taken turns in the spotlight. This time, Ant Lingbo released six models in one go, but the questions it wants to answer are more specific: Can large models trained in the digital world be directly installed into a robot’s body? Does the physical world require a whole new model system redesigned end-to-end—from perception and prediction to action?

Lingbo has chosen to rebuild this model system starting from the constraints of the physical world.

01 Why large models in the digital world can’t be plugged into a robot’s body

At the event, Shen Yujun shared a “get to the point” example.

Behind an opaque glass door there is a cat. A regular vision model can recognize the cat behind the door and can even describe the scene accurately. But when a robot needs to move toward the cat, “just seeing” is not enough. It must understand that the glass door forms a physical barrier—before the door is opened, the cat is in a space that the robot’s robotic arm cannot reach.

A digital model focuses on what’s in the scene; a robot must also judge distance, occlusion, contact relationships, and reachability. Semantic recognition being correct completes only the first step of the physical task.

Video models such as Imagination and Wanxiang serve content creation. Users provide a piece of text or a script; the model can refer to the full story, using more compute to buy better image quality and continuity.

Robots face time that only flows forward. When it grabs a cup, it doesn’t know whether someone will bump the table next second, nor whether the cup will slide. The model can only predict the next step based on the current state, and then adjust its actions when sensors return new information. Whether the picture looks good is not important; what matters is that prediction be reasonable, fast, and able to be converted into action.

The team calls this route “embodied-native,” and has trained LingBot-VA 2.0 from scratch. Public technical papers show that the model uses designs such as causal pretraining, sparse MoE, and asynchronous inference—aimed at high-frequency, closed-loop robot control.

This trade-off even allows the model to predict the scene with some distortions. When the robotic arm is preparing to pick up the cup, the generated cup from the model may not be sharp enough—as long as the direction of motion is correct. Sensors will continuously provide the real scene, and the model will then recalibrate based on the latest state.

VLA is easier to understand human intent in language and consumes fewer reasoning resources, making it a more implementable path today. Lingbo uses VLA to enter the scene, validate data, and then uses VA to explore dynamic modeling and future prediction. Shen Yujun believes that today’s separate technical routes each solve part of the puzzle, and in the future they may gradually merge into a single model.

Seen this way, Lingbo’s release of six models is more like dissecting single unresolved pain points in robot “brains.” The number of models may actually decrease in the future.

02 The primary cost of training from scratch is a data “long march”

Choosing an embodied-native approach immediately runs into a second problem: where does the data come from?

This question was asked again and again on site. Is one hundred thousand hours enough? Can a million hours produce an intelligence emergence moment? Can ten million hours bring robots to a ChatGPT-like moment?

Zhu Xing’s answer was direct: ten million hours may still not be enough.

Autonomous driving faces relatively clear traffic rules and driving tasks. A general robot needs to enter factories, warehouses, and homes; it must interact with objects of different materials, adapt to different bodies, and also handle failure states that cannot be defined in advance. The data distribution is far more complex than that of a single driving task.

According to public papers, LingBot-VLA 2.0’s pretraining data has grown from roughly 20k hours in the first generation to 60k hours, including 50k hours of robot trajectories and 10k hours of first-person human videos, covering 20 robot configurations across 17 manufacturers. The action space has also expanded from dual arms to include the head, waist, mobile base, and dexterous hands.

60k hours is still just the starting point. Lingbo places more emphasis on the speed and quality of the data closed loop.

Real data also includes human operation processes recorded via methods such as UMI and Ego, which can expand behavioral data at lower cost. The next stage also needs to add modalities such as touch and force sensation, and align them with first-person videos.

The team must continuously answer several engineering questions: Which data actually enters training? On what kinds of tasks does the model fail? Can new data collection tasks quickly cover capability gaps? How long does the whole pipeline—from collection, processing, and training to feedback—take?

As data scale grows, the team also needs to filter high-value samples. Autonomous driving has already gone through similar changes: early on, it chased volume; later, it finds the few samples that most improve the model from massive frames. For robots, anomalous and failure data is especially expensive, and it is even more likely to determine whether the model can handle long-tail issues.

Lingbo supports 20 configurations. Even after manufacturers integrate, further training still needs to be done around specific tasks. The role of pretraining is for the model to have seen different bodies in advance.

The real savings of one brain serving many machines is the cost you avoid by not training from scratch every time you change the platform body or add a new scenario.

03 The commercialization of robot “brains” first has to pass the success-rate test

A media outlet mentioned a warehousing case on site. When humans use a forklift, one搬运 task may take only 30 seconds, but a robot may take 1 minute or even longer—and if it encounters a new situation, it may stop to reassess.

Zhu Xing places success rate before speed. Even if a robot moves quickly, after it fails continuously a few times, the enterprise still needs staff to take over, and deployment becomes hard to create economic value. Once success rate is stable, companies can then compute cycle time, inference efficiency, and unit cost.

This is how the division of labor between base models and post-training is formed.

Zhu Xing compared pretraining to training a university student with excellent baseline skills. When the student enters a bank to work as an accountant, it still requires professional training. An embodied foundation model raises the ceiling of capability, while post-training turns the model into a production tool.

For robot manufacturers and scenario customers, post-training includes data collection, labeling, model adaptation, deployment, and inference optimization. Every link turns into cost. The smarter the base model is, and the more configurations and tasks it has seen, the fewer lessons post-training needs to add.

The commercial value of a general-purpose robot brain is lowering the investment required to develop models separately for each scenario. A factory screw-driving robot doesn’t need to learn how to wash dishes. Hotels and warehouses will also choose different bodies. Scenarios determine the body; a general brain needs to span more bodies.

Lingbo has already indicated it is pushing industrial deployment with body manufacturers and exploring different pricing methods such as outright purchase, subscriptions, and customization. However, on site it did not disclose customer cases that could be verified by outsiders, revenue scale, or cost models. What the market can confirm at this stage is the technical route and its ecosystem position; scaling into a closed-loop commercial model still requires waiting for more project data.

04 Why Lingbo is doing this big job

Training a robot brain from scratch requires long-term investment. Pretraining, data infrastructure, real-robot validation, and body adaptation—any one of these is hard for a small team to quickly fill in.

The core resources Ant provides to Lingbo include funding, talent, training infrastructure, data processing capabilities, and scenario ecosystem. On top of this foundation, Lingbo builds a full-stack model system—from spatial perception, video generation, and interactive world models to VLA and VA—and validates production readiness through collaborations with body partners.

This layout also reflects Ant’s judgment about the industrial landscape. Embodied intelligence is still at an early stage similar to a “battle of hundreds of models,” and in the future it may converge to only a few general foundation model providers. Robots are still far from entering homes at scale; it’s still too early to analogize to Windows or Android.

Looking at Ant Lingbo 2.0, model parameters and rankings are only part of it. More critical is whether it can continuously improve success rates across tasks, scenarios, and configurations, and whether it can reduce post-training costs to a level customers are willing to pay.

As Agents in the digital world rapidly spread with rising foundation-model capability, embodied intelligence may also experience similar “capability spillover.” The physical world adds one more unavoidable constraint: every judgment the model makes must ultimately be carried out by a real body.

Lingbo has chosen to rebuild this brain in advance. How far the roadmap can go ultimately depends on whether robots can truly get the job done.

Risk warning and disclaimer

        The market has risks; investment is需谨慎. This article does not constitute personal investment advice, and it does not consider the special investment objectives, financial conditions, or needs of any individual user. Users should consider whether any opinions, viewpoints, or conclusions in this article align with their specific circumstances. Investing based on this is at your own risk.