Raised hundreds of millions of yuan in investment from GL Ventures and other funds, LatentVerse: does not want to be labeled as a world model company, and is committed to building an embodiment-native brain | Exclusive debut by Intelligent Emergence

Text | Deng Yongyi

Editor | Zhang Yuxin

The embodied intelligence track may no longer need one more world model company — but a researcher born in 2001 still chose to launch his startup at this booming moment.

"There are too many companies developing world models in the market, and many of them have no unique characteristics of their own," said Hu Yucheng, founder of LatentVerse. However, he does not want to be classified into any of these companies, and what LatentVerse aims to build is not a world model either.

This 2001-born researcher is currently a PhD candidate at the Interdisciplinary Institute of Tsinghua University. In April this year, he left ByteDance Seed, and co-founded LatentVerse with members from institutions including Tsinghua University, Nanyang Technological University, Peking University, Qwen, ByteDance and Xiaomi, positioning the company to develop native embodied foundational models.

△ Hu Yucheng, CEO of LatentVerse

*Intelligence Emergence* has learned exclusively that embodied intelligence company LatentVerse has completed a seed round financing of several hundred million RMB, with the first batch of investors including GL Ventures, Qingliu Capital, Zhiyuan, StarDynamics, InnoAngel and other funds.

Over the past two years, VLA (Vision-Language-Action) has been the most mainstream technical route for embodied intelligence. From RT-2 to the popular Pi-0.5, the industry's default evolution path is that as long as the vision-language model is connected to action output and fed with sufficient robot operation data, the model can learn to act in the physical world.

But VLA has a fundamental limitation: it can only utilize data with action annotations. Such data is collected by manipulating robotic arms, with an average cost of more than 1,000 yuan per hour, and requires dedicated operators, leading to low efficiency.

Another route is the pure world model — it predicts future states through video generation and then maps them to actions, which lowers the data acquisition cost, as videos on the Internet can all be included in training. However, it cannot make use of general vision-language data, and lacks the semantic understanding and emergence capabilities brought by VLM. A model that only predicts picture changes cannot understand what the instruction "pass the cup to the guest" means.

LatentVerse has chosen the third route: instead of developing VLA or the currently popular WAM, it unifies VLM and the world model into the same architecture.

In fact, Hu Yucheng was one of the first researchers to prove that "world models can directly guide robot actions". In early 2024, he proposed the multimodal diffusion VLA (PAD), and in November of the same year, he proposed VPP, the first video action model to realize closed-loop reasoning. The latter is known in the industry as the pioneering work of WAM (World Action Model), and later NVIDIA developed DreamZero on this basis.

After joining ByteDance Seed, he led the R&D of the BagelVLA series of embodied foundational models, and the publication time of related work was even earlier than Pi07, the representative work of Physical Intelligence.

His experience in the industry made Hu Yucheng see the arrival of an important technical inflection point: to reach the end point of physical AGI, the key is to solve the data bottleneck — neither the pure VLA route nor the pure world model route can reach the end, and a foundational model that unifies understanding, prediction and action in the same architecture is required.

With the reduction of data collection costs, entering 2026, data has become the most important issue for embodied intelligence, which has gradually become a consensus in the embodied intelligence field.

Bet on the Unified Tactile Action Model

The first model LatentVerse is training, UTAM (Unified Tactile Action Model), is evolved from the UAM (Unified Action Model) previously proposed by the team, and outputs visual, language and tactile feedback signals at the same time.

More specifically, the model contains four expert modules inside, forming a complete execution chain: the vision-language expert is responsible for understanding task intentions; the world model expert combines the skills learned from past videos to plan a new task flow; the action expert generates a coarse-grained action representation (Latent action) from the reasoning chains of the first two, which is equivalent to a large action instruction such as "reach out to pick up the cup"; finally, the high-frequency reasoning tactile expert decodes this coarse action into fine end control signals — for example, adjusting the force and posture in real time after the finger touches the cup wall. The first half is open-loop execution, and the second half is tactile closed-loop correction, which is similar to the human operation logic of "reach out first, then fine-tune by feel".

The choice of model route is derived from the scenarios that the team bets on in reverse. "Generalization, long-horizon performance and dexterity are not three independent technical indicators, but capability conditions that robots must meet at the same time for commercialization," said Hu Yucheng. "We believe that the embodied foundational model should have stronger generalized capabilities, such as in home scenarios."

The technical route that LatentVerse bets on involves more data collection dimensions and longer-horizon tasks. For example, robots need to clean hotel rooms, pack drinks in narrow spaces, grab soft objects in homes, and so on.

In LatentVerse's framework, robots need to have two capabilities at the same time when entering real scenarios: cognitive intelligence and contact intelligence. Cognitive intelligence is responsible for understanding instructions and planning steps — which is what VLA and world models are already doing.

But cognition alone is not enough. The contact information of the last centimeter determines whether the robot can actually complete the operation. For example, whether the surface of the water cup is smooth, whether deformation occurs when a flexible object is grabbed, and other such information cannot be obtained purely through vision, and must be corrected in real time through tactile feedback.

These tasks require not only predicting what will happen in the next frame, but also having the capabilities of understanding instructions, planning steps, predicting physical changes, perceiving contact and generating precise actions at the same time.

In the past two years, the embodied intelligence track could only achieve some precise operations in fixed scenarios such as industry, but it is still very difficult to achieve capability generalization on routes such as VLA.

The advantage of the unified embodied model route is that the range of data that can be utilized is much larger than that of pure VLA. As long as videos related to operations, cross-robot machine data, and general VLA data can all be incorporated into the training framework, when the data scale increases, the model's generalization ability for scenarios and objects will be significantly improved.

But the difficulties are also obvious: there is insufficient high-quality data available. Even in the past, the development of the embodied intelligence track has rapidly expanded the scale of data collection, but most of these data are collected for specific tasks and are difficult to reuse in the training of general task capabilities.

Hu Yucheng said that to achieve capability leap on the same model route, "first, the data scale needs to be increased, and second, the model must be able to utilize massive data — not only the data collected by itself, but also Internet-scale video data and cross-robot heterogeneous data. The data processing dimension is also more complex, which requires describing how to decompose the entire task and applying it in different training stages."

For this reason, in the few months since its establishment, LatentVerse has built a complete link covering data collection, data processing and model training, which first includes a self-developed data collection pipeline, as well as the embodied native world model that LatentVerse is training.

From the deployment perspective, native embodied models also have structural advantages.

At present, embodied intelligence generally faces the problem of "good demo performance but difficult to deploy on a large scale". The core reason is the insufficient robustness of the model — changes in light, human interference, and inconsistent scene layout in real scenarios will lead to a much higher failure rate than in demos. "VLM experts can generate human intentions, understand physical information and make next-step judgments. World model experts can consume massive amounts of data to improve the robustness of scenarios, objects and tasks, and convert discrete texts into semantically rich visual signals," said Hu Yucheng.

More technically challenging than VLA, covering wider scenarios than factories

The premise for the native embodied model route to work is data.

Choosing this training route comes at the cost of a sharp increase in training difficulty: in data processing, long-horizon tasks need to be decomposed into several steps and annotated step by step, which is far more complex than processing images or texts; training requires multi-stage and multi-task execution at the same time; the WAM route relies on diffusion models, and reasoning latency, for example, is an engineering problem that has not been fully solved yet.

"Our team has worked on the pure VLA route and the pure world model, so we know the limitations of each," said Hu Yucheng, founder. Another example is that another co-founder of LatentVerse is responsible for dexterous hand algorithm and motion control, which can fill the shortcomings on the control side — after the data link from the robot body to the human hand is connected, LatentVerse hopes to build a data flywheel for embodied scenarios.

To this end, LatentVerse has also formed a young and refined team. The algorithm team covers sub-fields related to robots such as core embodied intelligence algorithm modules, dexterous hands, simulation, and 3D vision.

At present, LatentVerse's data strategy is divided into three layers: manipulation data (VR collection with precise action annotations) is only used to train the action expert; human hand collected data (such as Ego-centric) and open source video data are used to train the understanding and world model experts.

On the data collection side, LatentVerse bets on the robot-agnostic data collection route. The team is also developing a self-designed data collection glove, and has established initial cooperation with several data collection bases. "Existing data collection gloves are not sufficient to meet our requirements," Hu Yucheng said. In the future, through wearable devices, non-professionals can participate in data collection, and even integrate it into daily life. LatentVerse's data collection solution reduces the per-hour collection cost from the 1000-yuan level to less than 100 yuan.

However, what determines the model's capability is more about data quality and how the data is used.

Another uniqueness of the LatentVerse team is that the core members of the team have been working on the unified model route since the beginning of 2024, and most of their work is pioneering rather than following others, and has become an important technical foundation of the industry.

For example, Zhang Jianke, another co-founder of LatentVerse, is the leader of many pioneering works in the embodied intelligence field — from HiRT (the fast-slow system for embodied intelligence, which is followed by Figure's Helix) to UP-VLA (the first generation-integrated embodied model), and then to VLM4VLA, which systematically proves the inherent limitations of VLM as an embodied foundation model.

"The capability emergence of embodied models cannot simply copy the development path of LLM. While promoting large-scale data, we should also make full use of the structured information in various heterogeneous data to build a unified representation," said Hu Yucheng.

According to LatentVerse's plan, the company will release its first 16B parameter embodied foundational model in one quarter, and accumulate hundreds of thousands of hours of training data within one to one and a half years.

In terms of scenario selection, LatentVerse hopes to incorporate more complex dimensional signals such as tactile into the model, and after achieving sufficient generality and generalization capabilities, cover a wider range of industrial, commercial, and household scenarios.

This comes from the LatentVerse team's different understanding of world models. In his definition, any model that predicts the next token can be regarded as a world model, and the real route difference among different players lies in how to define the state space. "We believe we are a native embodied world model, a world model for robotics."

Behind this definition is actually a judgment of scenarios. The main reason why embodied intelligence can only be applied in vertical scenarios at present is the insufficient capability of the "brain". Only when generality and generalization capabilities are fully achieved can it enter unstructured scenarios such as homes.

"What embodied intelligence needs to cross is the last centimeter between cognition and contact." In Hu Yucheng's view, robots should not only learn to see the world, but also learn to touch the world. First of all, robots need to learn to provide practical value in the real world.

Image source | AI generated

Welcome for communication

Welcome to follow us

This article is from the WeChat official account "Intelligence Emergence", author: Deng Yongyi, published by 36Kr with authorization.