Throughout the communication session, the media repeatedly posed the same core question in various forms.
How far is today's robotic brain from being truly capable of working in the physical world?
The assessment given by Ant Lingbo's CEO Zhu Xing and Chief Scientist Shen Yujun was considerably more measured than the current market fervor. The current robotic brain may not yet have reached its "GPT-1 moment." The industry has not yet witnessed a true emergence of intelligence, and technological pathways are far from converging.
Over the past year, concepts like VLA, world models, and video action models have taken turns in the spotlight. Ant Lingbo's release of six models at once aims to answer a more specific question. Can large models trained in the digital world be directly installed into a robot's body? Does the physical world require a completely redesigned model system, from perception and prediction to action?
Lingbo has chosen to start from the constraints of the physical world and rebuild this model system from scratch.
Digital World Models and Robotic Physical Limitations
At the event, Shen Yujun illustrated the challenge with an example of "opening a door to see a cat."
Imagine a cat behind an opaque glass door. A standard vision model can identify the cat and accurately describe the scene. However, for a robot intending to move toward the cat, merely "seeing" is insufficient. It needs to understand that the glass door presents a physical barrier; until the door opens, the cat occupies a space unreachable by the robotic arm.
While digital models focus on what is present in an image, a robot must also judge distance, occlusion, contact relationships, and reachability. Correct semantic recognition is only the first step in completing a physical task.
Models like Sora and Wanxiang serve content creation. Given a text prompt or script, they can reference a complete narrative, allocating more computational power for higher visual quality and continuity.
In contrast, a robot operates in a timeline that only moves forward. When reaching for a cup, it doesn't know if someone will bump the table the next second or if the cup will slide. The model can only predict the next action based on the current state and then correct its movements as new sensor data arrives. Visual beauty is irrelevant; predictions must be reasonable, fast, and convertible into action.
The team terms this approach "embodiment-native" and has trained LingBot-VA 2.0 from the ground up. Published technical papers indicate the model employs designs like causal pre-training, sparse Mixture-of-Experts (MoE), and asynchronous inference, all serving high-frequency, closed-loop robotic control.
This trade-off even allows for some distortion in predicted images. When a robotic arm prepares to grasp a cup, the model-generated image of the cup need not be perfectly clear, as long as the movement direction is correct. Sensors continuously provide real-world visuals, allowing the model to recalibrate based on the latest state.
VLA models, which more easily understand human language intent and consume fewer inference resources, represent a more immediately viable path. Lingbo uses VLA to enter scenarios and validate data, then employs VA to explore dynamic modeling and future prediction. Shen Yujun believes today's distinct technical paths are each solving a piece of the puzzle, potentially converging into a single model in the future.
From this perspective, Lingbo's release of six models appears more like an effort to deconstruct the still-unresolved point problems of the robotic brain. The number of models may actually decrease in the future.
The Primary Cost of Ground-Up Training: A Data Long March
Choosing an embodiment-native path immediately raises a second question: where does the data come from?
This issue was repeatedly pressed at the event. Is ten thousand hours enough? Could a million hours trigger an intelligence emergence? Would ten million hours bring about robotics' "ChatGPT moment"?
Zhu Xing's response was direct: even ten million hours might not be sufficient.
Autonomous driving deals with relatively clear traffic rules and driving tasks. A general-purpose robot needs to enter factories, warehouses, and homes, interact with objects of different materials, adapt to different physical embodiments ("bodies"), and handle undefined failure states. Its data distribution is far more complex than a single driving task.
Published papers show that the pre-training data for LingBot-VLA 2.0 has increased from about 20,000 hours in the first generation to 60,000 hours. This includes 50,000 hours of robot trajectory data and 10,000 hours of first-person human video, covering 20 different robot configurations from 17 manufacturers. The action space has also expanded from dual arms to include the head, waist, mobile base, and dexterous hands.
Sixty thousand hours is still just a starting point. Lingbo places greater emphasis on the speed and quality of the data feedback loop.
Real-world data also includes human operation processes recorded via methods like UMI and Ego, allowing for lower-cost expansion of behavioral data. The next phase will involve adding modalities like tactile and force sensing, aligned with first-person video.
The team must continuously address several engineering questions. Which data actually enters training? On which types of tasks does the model fail? Can new data collection tasks quickly cover capability gaps? How long is the cycle from collection, processing, and training to feedback?
As data scales, the team must also filter for high-value samples. Autonomous driving has undergone a similar evolution, initially pursuing volume before later identifying the few frames from massive datasets that most improve the model. Data on robot anomalies and failures is particularly expensive and more likely to determine if a model can handle long-tail problems.
While Lingbo supports 20 configurations, manufacturers still need to conduct task-specific fine-tuning after integration. The role of pre-training is to expose the model to different physical embodiments in advance.
The true efficiency gain of a "one-brain, many-machines" approach lies in saving the cost of training from zero each time a new robot body is introduced or a new scenario is added.
The Commercialization Hurdle: Prioritizing Success Rate
At the event, a media representative cited a warehouse case. A human using a forklift might complete a搬运 task in 30 seconds, while a robot could take a minute or longer, potentially stopping to re-evaluate when encountering a novel situation.
Zhu Xing prioritizes success rate over speed. No matter how fast a robot acts, if it fails several times in a row, a human still needs to take over, making deployment difficult to justify economically. Only after achieving a stable success rate will companies further calculate cycle time, inference efficiency, and unit cost.
This creates a division of labor between the foundation model and subsequent fine-tuning.
Zhu Xing compares pre-training to educating a university student with strong foundational qualities. That student still needs professional training to become an accountant at a bank. The embodied foundation model raises the ceiling of capability, while fine-tuning transforms the model into a production tool.
For robot manufacturers and end customers, fine-tuning encompasses data collection, annotation, model adaptation, deployment, and inference optimization. Each step translates into cost. The smarter the foundation model and the more configurations and tasks it has encountered, the less remedial work is needed during fine-tuning.
The commercial value of a general-purpose robotic brain lies in reducing the investment required to develop a model separately for each scenario. A robot screwing bolts in a factory doesn't need to learn to wash dishes; hotels and warehouses will choose different robot bodies. The scenario dictates the body, so a general-purpose brain must span more embodiments.
Lingbo has stated it is advancing industrial deployment with hardware manufacturers and exploring various monetization models like outright purchase, subscription, and customization. However, the event did not disclose customer cases, revenue scale, or cost models for external verification. At this stage, the market can primarily confirm the technical roadmap and ecosystem positioning; a scaled commercial闭环 still awaits more project data.
The Rationale for Undertaking This Heavy Lift
Training a robotic brain from the ground up requires long-term investment. Pre-training, data infrastructure, real-world validation, and hardware adaptation—each presents a significant challenge difficult for a small team to overcome quickly.
The core resources Ant provides to Lingbo include funding, talent, training infrastructure, data processing capabilities, and access to scenario ecosystems. On this foundation, Lingbo is building a full-stack model system encompassing spatial perception, video generation, interactive world models, VLA, and VA, then validating deployment capabilities through partnerships with hardware manufacturers.
This strategic layout also reflects Ant's assessment of the industry landscape. Embodied intelligence remains in an early stage akin to the "hundred-model war," potentially converging into a few providers of general-purpose foundation models. Robots are still far from entering households on a large scale; it's too early to draw analogies to Windows or Android.
Observing Ant Lingbo 2.0, model parameters and benchmark rankings are only part of the story. More critical metrics are whether it can continuously improve success rates across tasks, scenarios, and configurations, and whether it can reduce fine-tuning costs to a level customers are willing to pay.
Just as digital world agents proliferated rapidly after foundational model capabilities advanced, embodied intelligence may experience a similar capability溢出. However, the physical world adds an inescapable layer of constraint: every judgment the model makes must ultimately be executed by a real, physical body.
Ant Lingbo has chosen to rebuild this brain in advance. How far this path leads will ultimately depend on whether the robots can truly get the job done.
Comments