The embodied intelligence sector may not need another world model startup. Hu Yucheng, a researcher born in 2001, founded LatentVerse anyway, even as he insists the company should not be grouped with the growing field of world model developers.
“There are too many companies working on world models, and many of them have no distinguishing features of their own,” LatentVerse founder Yucheng Hu said. But he does not want LatentVerse grouped with any of them. Nor is a conventional world model what the company is trying to build.
Hu is a doctoral student at Tsinghua University’s Institute for Interdisciplinary Information Sciences. In April, he left ByteDance Seed and co-founded LatentVerse with members from Tsinghua University, Nanyang Technological University, Peking University, Qwen, ByteDance, Xiaomi, and other organizations. The company is focused on what it calls embodied-native foundation models.
According to 36Kr, LatentVerse has completed a seed funding round worth a nine-figure RMB sum. Investors include GL Ventures, Crystal Stream, Agibot, Robot Era, and Innoangel Fund.
Over the past two years, VLA, or vision-language-action, has emerged as a mainstream technical approach to embodied intelligence. From RT-2 to the widely discussed Pi-0.5, the development path has been relatively straightforward: connect a vision-language model, or VLM, to an action output, train it on enough robot manipulation data, and have the model learn to act in the physical world.
One limitation of VLA systems is their reliance on data with action labels. Collecting such data through teleoperated robotic arms costs more than RMB 1,000 (USD 148.1) per hour on average, requires specialized teleoperators, and can be inefficient.
Another approach is the pure world model, which predicts future states through video generation and then maps those predictions into actions. Data can be cheaper to acquire because internet videos can also be incorporated into training. But this approach does not fully make use of general-purpose vision-language data, limiting the semantic understanding and emergent capabilities that VLMs can provide. A model trained only to predict changes in an image, for example, cannot necessarily understand what an instruction such as “hand the cup to the guest” means.
LatentVerse has chosen a third path. Rather than building either a VLA model or today’s increasingly popular world action model, or WAM, it is seeking to unify VLMs and world models within the same architecture.
Hu’s previous research has explored how world models can guide robot actions. In early 2024, he proposed PAD, a multimodal diffusion VLA approach. In November that year, he introduced VPP, a video-action model designed for closed-loop inference. The work has been described as foundational to later WAM development, with Nvidia subsequently developing DreamZero on that foundation.
After joining ByteDance Seed, Hu led the development of the BagelVLA series of embodied foundation models. The related work was published before Physical Intelligence’s Pi07.
His experience convinced him that embodied intelligence had reached an inflection point. In Hu’s view, reaching physical artificial general intelligence, or AGI, requires overcoming the sector’s data bottleneck. Neither a pure VLA nor a pure world model can do that alone, he believes. Instead, embodied systems need a foundation model that unifies understanding, prediction, and action within a single architecture.
As data collection costs fall, the challenge is increasingly shifting toward obtaining enough usable data and building models that can learn effectively from it.
A unified tactile action model
LatentVerse is now training its first model, UTAM, short for unified tactile action model. Building on the team’s earlier UAM, or unified action model, UTAM is designed to output visual, language, and tactile signals simultaneously.
The model contains four expert modules that together form an execution chain. A vision-language expert interprets the intent behind a task. A world model expert draws on skills previously learned from video to plan a new task sequence. An action expert then uses outputs from the first two modules to generate a coarse action representation, or latent action. This could correspond to a broad instruction such as reaching out and picking up a cup.
Finally, a tactile expert operating at a high inference frequency translates that coarse action into precise end-effector control signals. It adjusts force and posture in real time when, for example, a finger touches the side of a cup.
The first part of the process relies on open-loop execution, while the second uses closed-loop tactile correction. The idea resembles how a person first reaches toward an object, then fine-tunes the movement based on touch.
The architecture reflects the scenarios LatentVerse is targeting.
“Generalization, long-horizon execution, and dexterity are not three independent technical metrics. They are capabilities that robots must possess simultaneously if they are to be commercialized,” Hu said. “We believe embodied foundation models should have stronger general-purpose and generalization capabilities, including in settings such as the home.”
LatentVerse’s technical approach involves collecting data across more dimensions and training for longer-horizon tasks. Target scenarios include robots cleaning hotel rooms, packing beverages in cramped spaces, and grasping soft objects in homes.
Under LatentVerse’s framework, a robot operating in the real world needs two capabilities at once: cognitive intelligence and contact intelligence. Cognitive intelligence covers understanding instructions and planning steps, tasks that VLA and world models already attempt to handle.
But cognition alone is not enough. Information from the final centimeter of physical contact can determine whether a robot completes an operation. A robot may need to determine, for example, whether the surface of a drinking glass is slippery or whether a soft object deforms when grasped. Vision alone cannot capture all of this information, so the robot must make real-time corrections using tactile feedback.
These tasks require more than predicting what happens in the next frame. A robot must understand instructions, plan steps, predict physical changes, sense contact, and generate precise actions.
Embodied intelligence systems have already demonstrated capabilities in some fixed scenarios, including certain precision operations in industrial settings. Achieving broader generalization through approaches such as VLA, however, remains difficult.
One advantage LatentVerse sees in a unified approach is access to a broader range of training data than pure VLA systems can use. Manipulation videos, cross-embodiment robot data, and general-purpose VLA data can all be incorporated into the training framework. LatentVerse expects generalization across environments and objects to improve as the scale and diversity of that data increase.
The constraint is equally clear: there is still not enough usable, high-quality data. Although embodied intelligence development has expanded data collection in recent years, much of that data was gathered for specific tasks and can be difficult to reuse when training general-purpose capabilities.
Hu said achieving a major jump in capability along the same model path requires two things.
“First, the scale of the data has to increase. Second, the model has to be able to use massive amounts of data, not just what we collect ourselves, but internet-scale video data and heterogeneous data across different embodiments,” he said. “The dimensions involved in data processing also become more complex. You need to describe how an entire task is broken down and use that information at different stages of training.”
LatentVerse said it has spent the months since its founding building a pipeline spanning data collection, processing, and model training. This includes a proprietary data collection system and the embodied-native world model the company is now training.
The company also sees potential advantages at deployment.
Embodied intelligence systems can perform well in demonstrations but prove harder to deploy at scale, largely because their models are not robust enough for changing real-world conditions. Variations in lighting, human interference, and differences in how environments are arranged can increase failure rates outside controlled demonstrations.
“The VLM expert can infer human intent, understand physical information, and make a judgment about what to do next. The world model expert can ingest massive amounts of data to improve robustness across environments, objects, and tasks, while turning discrete text into semantically rich visual signals,” Hu said.
A harder training path beyond factory settings
Data is a prerequisite for making an embodied-native model work, but LatentVerse’s approach also increases training complexity.
Processing long-horizon tasks requires breaking them into multiple steps and annotating those steps sequentially, making the process more complex than conventional image or text processing. Training must also span multiple stages and tasks. Some WAM approaches rely on diffusion models, while inference latency remains an engineering challenge.
“Our team has worked on pure VLA approaches and pure world models, so we know where the limitations of each lie,” founder Hu said.
Another LatentVerse co-founder focuses on dexterous-hand algorithms and motion control, adding expertise on the control side. Once a data pipeline linking robot embodiments with human hand data is established, the company hopes to create a self-reinforcing data loop for embodied intelligence applications.
LatentVerse has assembled a small team specializing in embodied intelligence algorithms, dexterous hands, simulation, 3D vision, and other areas of robotics.
Its data strategy spans three categories. Teleoperation data, collected using virtual reality and carrying precise action labels, is used only to train the action expert. Human hand data, including egocentric data captured from a person’s point of view, is used to train the understanding and world model experts. Open-source video data is used for the same purpose.
For data collection, LatentVerse is betting on an embodiment-agnostic approach, meaning data collection methods that are not tied to a specific robot form. The team is developing its own data collection glove and says it has established preliminary partnerships with several data collection facilities.
“Existing data collection gloves still do not adequately meet our requirements,” Hu said.
LatentVerse eventually plans to use wearable devices to enable nonprofessionals to contribute data, potentially integrating collection into everyday activities. The company estimates that this approach could cut collection costs from more than RMB 1,000 per hour to less than RMB 100 (USD 14.8) per hour.
For Hu, however, model performance ultimately depends on both data quality and how that data is used.
LatentVerse’s core members have been working on unified models since early 2024. Co-founder Zhang Jianke, for example, has led several embodied intelligence projects. They include HiRT, short for hierarchical robot transformer, a fast-slow system that was adopted by Figure’s Helix; UP-VLA, described as the first embodied model to unify generation and understanding; and VLM4VLA, which examined the limitations of VLMs as embodied foundation models.
“The emergence of capabilities in embodied models cannot simply copy the development path of large language models,” Hu said. “While scaling data, we should also fully use the structured information contained in different types of heterogeneous data to build unified representations.”
Under LatentVerse’s plan, the company will release its first 16-billion-parameter embodied foundation model within one quarter. Over the next 1–1.5 years, it plans to accumulate hundreds of thousands of hours of training data.
LatentVerse also wants to incorporate more complex signal dimensions, including touch, into its models. If it can develop stronger general-purpose and generalization capabilities, it aims to address a broader range of industrial, commercial, and household scenarios.
That strategy reflects Hu’s broad definition of a world model. In his view, any model that predicts the next token can be considered a world model. The meaningful distinction lies in how each approach defines its state space.
“We believe ours is an embodied-native world model, a world model for robotics,” he said.
Behind that definition is a view of where embodied intelligence can ultimately be used. Hu argues that one reason robots remain concentrated in narrow, vertical-specific scenarios is that their underlying intelligence is not capable enough. Only when general-purpose capabilities and generalization improve sufficiently, he believes, can robots operate reliably in unstructured environments such as homes.
“The last centimeter embodied intelligence must cross is the gap between cognition and contact,” Hu said.
Robots, in his view, cannot just learn to understand the world. They also need to learn to touch it, and to turn those capabilities into practical value in real-world environments.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Deng Yongyi for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.75 = USD 1 based on estimates as of August 18, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.