Harnessing Human Data: Lingchu Intelligence’s Approach to Embodied AI in China

Global Perspective on China: Lingchu Intelligent Utilizes 100,000 Hours of Human Data to Craft an Embodied Intelligence Answer

As of May 14, 2026, the term “world model” has become increasingly prevalent in the embodied intelligence industry. Numerous companies are now framing their technological direction around robotic world models, aiming to enhance robot training efficiency through learnable environmental models. Lingchu Intelligent (PsiBot) is often included in this narrative. However, co-founder Chen Yuanpei believes that the world model is not the core focus of Lingchu, but rather a tool to facilitate data transfer. “I don’t see developing a world model as a transformation. It’s merely a tool. From day one, we have been focused on human data,” he states. The primary question for Lingchu is whether real human operational data can be scaled into training data for robots.

Before Lingchu was established, Chen Yuanpei had already begun exploring the use of human hand movement data to train for dexterous operations. This work, later published in CoRL 2024, became a key technological source for Lingchu’s commitment to the human data route. Currently, Lingchu has reached a clearer conclusion through large-scale data practices: at the level of 100,000 hours, human data can significantly replace data collected from actual machines. In this approach, VLA, world models, reinforcement learning, and exoskeleton gloves are not isolated endpoints; instead, they collectively aim to construct a pipeline that translates human data into robot policies.

According to statistics from Morgan Stanley Research, by the end of April 2026, global venture capital financing for humanoid robots in 2026 had already surpassed the total for 2025. With capital accelerating into the sector, the data, models, and implementation capabilities within the embodied intelligence industry are being repriced.

1. A Human Data Route from Day One: Shifting from Robot-Centric to Human-Centric

The data issue in embodied intelligence fundamentally revolves around balancing scale, quality, and transfer efficiency. Over the past few years, one mainstream approach has been teleoperation, where humans directly control robots or shadow arms using similar or nearly identical remote operation devices. This type of data is closely aligned with the robot itself, making transfer easier and the training process more straightforward. However, teleoperation has its own clear drawbacks: high collection costs, heavy equipment, strong site dependency, and the need for trained operators, making it difficult to gather a sufficiently large data set. Companies aiming to train general robotic capabilities quickly encounter limits with this data production method.

Another approach is ego data, which involves using cameras to capture human first-person perspective operating data. This method is less costly and more closely resembles actual human behavior, but it introduces new challenges. There is a natural gap between humans and robots: differences in joint structure, degrees of freedom, motion habits, and visual perspectives complicate the direct application of this data for robot training, often resulting in low transfer efficiency, high noise, and non-standard actions. Lingchu’s conclusion is that while transfer issues can be addressed through models and algorithm pipelines, data scale challenges must be resolved through collection methods. “One core reason we designed this glove was to minimize the impact on a person’s daily tasks. For instance, when a cashier wears our glove, it doesn’t significantly affect their work, whereas using two robotic claws would hinder basic tasks like scanning,” Chen explains.

This means that Lingchu aims to enter real labor scenarios—such as logistics, warehousing, cashiering, and manufacturing—where human operational behaviors are continuously generated, as opposed to specially constructed robotic data production environments. This approach contrasts with robot-centric solutions like UMI, which collect data using devices more closely resembling robotic claws. While robot-centric methods offer higher transfer efficiency, they limit operator movements and struggle to access real labor scenarios. Lingchu has chosen a human-centric path, accepting greater transfer difficulty in exchange for a higher data scale limit. Currently, Lingchu is simultaneously collecting two types of human data. The first is exoskeleton glove data, which captures hand and arm movements through mechanical connections without relying on IMUs, offering higher precision and the ability to record a complete range of motion. The second is purely visual data, which utilizes head and wrist cameras to document human operations without gloves, resulting in lower costs and stronger scalability, though with relatively weaker motion precision.

In Lingchu’s design, the glove’s degrees of freedom are maximized not only to adapt to their dexterous hand but also to enhance the cross-body transferability of data. In other words, Lingchu aims to collect data that is not tied to a specific robotic hardware but can be transferred to different robotic entities in the future.

2. W0 and R2: Using World Models to Complete the Transfer from Human Data to Robot Policies

The core difficulty of the human data route lies not in collection but in transfer. Human actions inherently contain noise and irregularities, and human dynamics differ from those of robots. Lingchu’s solution is to use reinforcement learning to achieve transfer within a world model. Lingchu’s system consists of two main modules: W0 and R2. R2 represents the policy that is ultimately deployed on the robot, responsible for executing operations. W0, on the other hand, is a world model—more accurately, an action-conditioned world model that predicts the next state given the current state and action. During the training phase, W0 acts as a learnable simulator, while R2 iteratively refines itself through online reinforcement learning in the environment constructed by W0. W0 provides environmental feedback, and R2 explores within it, transferring human dynamics to robot dynamics and generating new training data to refine R2, thereby forming a closed loop. During deployment, W0 is no longer in use, with only R2 running on the robot. “Once the model is well-trained, there’s no need for W0 anymore. W0 is a simulator; it’s part of the improvement process, not part of the deployment,” Chen notes. This is also why he hesitates to simply define Lingchu as a “world model company.” In his view, the world model is not a standalone direction but an intermediate module in the data transformation pipeline. Its purpose is not to replace the real world but to assist in transferring human data into robot policies.

“The algorithms themselves aren’t that important; we use whichever is most effective. Our core focus remains human data and the pipeline that transforms this data into high-quality robot data,” he emphasizes. In this system, data quality is not solely reliant on manual review. Lingchu entrusts the evaluation to the model itself: whether a piece of data can successfully convert within the world model and allow the policy to function is the selection criterion. Data that passes the test is retained, while data that does not is discarded. As the model’s capabilities improve, the boundary for data selection will also change dynamically. Chen believes that a significant milestone for Lingchu’s human data route is the internal validation achieved at the level of 100,000 hours. “We have very little actual machine material; genuine machine data is extremely scarce. Yet, we can achieve results comparable to companies that have gathered tens of thousands of hours of teleoperation data using human data alone,” he states. This raises a fundamental question: Is it necessary for a robot’s foundational model to rely on large-scale real machine teleoperation data? Chen’s assessment is that while real machine data remains important, it is not the only fuel needed. If human data collection is sufficiently scalable and the transfer pipeline is effective, then a substantial amount of real machine data can be partially replaced by human data. Real machine data becomes more like a supplement for calibration, validation, and minor fine-tuning, rather than the sole source of data. This does not imply that human data is inherently equivalent to robot data. On the contrary, for human data to be genuinely usable, it must undergo a complete pipeline involving collection systems, world models, reinforcement learning, data filtering, and policy training. Lingchu aims to establish this system capability.

3. From Dataset to Implementation: SynData, Small Stack, and Route Boundaries

As of May 13, 2026, the SynData dataset from Lingchu Intelligent has been downloaded approximately 14,600 times on Hugging Face. This new generation of large-scale, real-world multimodal datasets is based on the R2 and W0 systems, covering dimensions such as vision, language, and actions. Leveraging their self-developed exoskeleton glove system, SynData can capture high-precision operational data of hands and arms with complete degrees of freedom, while also incorporating bare hand data and natural human interaction behaviors for open use in action modeling, operational learning, prop learning, and multimodal intelligent research.

For Lingchu, SynData represents a phase of externalization of its technological route: based on real human operational data, it achieves transfer through world models and reinforcement learning, ultimately training deployable policies for robots. However, from a commercialization perspective, Chen does not believe the industry has entered the stage of “general base models” yet. He categorizes Lingchu’s current position into several layers. The first layer is the production capacity phase, where the primary source of revenue still comes from hardware, including exoskeleton gloves, collection systems, and material field construction. Data revenue is expected to gradually become the mainstay by next year. The second layer is the policy adjustment phase, where robots entering specific customer scenarios still require policy adjustments based on task, environmental, and timing requirements. “A truly general base model that doesn’t need adjustments will take around three to five more years,” he estimates. The third layer is the base model phase, which is the goal but not the current state. This also clarifies why Lingchu has opted for a “small stack” approach. Chen defines a “small stack” as focusing on models while controlling key processes but stopping at core components. For example, components like tactile sensors and precision reducers are sourced externally rather than developed in-house, not to showcase full-stack capability but to facilitate implementation. “To achieve implementation, hardware stability and timing requirements are highly coupled with hardware, and currently, it’s not feasible; we have no choice but to develop it ourselves,” he notes.

At this stage, implementing robots is not merely a software issue. The stability of a policy depends on the robot’s body, actuators, sensors, control systems, task timing, and scene constraints. Models and hardware remain highly coupled, making it challenging to achieve real-world delivery by focusing solely on models. Regarding other industry routes, Chen has clear judgments. Concerning the recent attention on robot demos like Genesis, he believes they should neither be glorified nor dismissed. “If we trained on that hardware, we could also replicate their demos,” he asserts. On the topic of simulation, he expresses relative pessimism. He views simulation as an important tool but believes that expecting a significant breakthrough in simulation to independently solve issues of contact, long tails, and high-precision operations in the real physical world is unlikely.

So, could the human data route be invalidated? Chen posits that if this route were to be disproven, it would likely be due to one of two possibilities: first, a significant breakthrough in simulation that can generate sufficiently realistic, diverse, and transferable data at low cost; or second, a company possessing enough financial and engineering capabilities to genuinely operate the real machine data flywheel. He considers the latter scenario to be more probable. In his view, the real challenge for the human data route is not to prove that human data is cleaner than real machine data, but to demonstrate whether a superior comprehensive solution can be achieved across scale, cost, transfer efficiency, and generalization capability. The competitive advantage of this route extends beyond mere data volume. “Algorithms don’t hold any secrets. However, data—including the entire data organization pipeline, accumulation, and processing methods—will affect you for a long time. Some people may have spent three years accumulating data; catching up to them immediately is very difficult,” he states. Beyond data, organizational capability is also crucial. “The entire organizational culture, structure, and values are also very important,” he adds.

From research papers to company formation, Chen believes that Lingchu has consistently pursued the same goal: to enable robots to utilize human data effectively. VLA, world models, and reinforcement learning are merely tools; the true direction is the development of general robotic capabilities.