Huang Xiaoyi | H997HbiuYoky | yokyliu617 At this year's WRC, the "concept hype" is fading. VLA, repeatedly showcased at booths, in keynote speeches and interviews last year, is rarely demonstrated as an end-to-end complete solution this year, and "end-to-end" has also taken a back seat. The new dominant buzzword — "world model" — still lacks a unified, definitive definition: some generate future videos, some predict events, and others simulate in latent space. However, the debates over technical routes seen last year have disappeared.
More people are realizing: All debates over algorithmic routes lacking data support ultimately prove to be illusions.
Whether developing VLA, world models, or end-to-end embodied foundation models, it all comes back to data. Without sufficient real-world physical interaction data, VLA merely connects vision, language, and action; end-to-end simply hides module boundaries within the model; and world models struggle to truly understand how the world changes after an action is taken.
Therefore, this year, the question has shifted from "What is the ultimate route for the robot brain?" to two more specific issues: What kind of data should be collected, and how should it be collected?
We spoke with six key companies in the industry and referenced their public speeches to compile four core observations.
I. Is VLA Really Obsolete? Last year, the industry was still debating whether VLA was the answer for the robot brain. This year, the more accurate question has become: To what extent has VLA actually been "eliminated"? Based on responses from multiple companies, VLA has not disappeared, but it is transitioning from an independent technical route to a component within a system. What is truly obsolete is the illusion that "VLA alone is enough."
Keenon Robotics: VLA is the Skeleton, World Models Enable It to "Think Before Acting" Last year, the industry indeed had extensive discussions on routes like end-to-end, VLA, and world models, each with its own supporters. However, from our practical experience, some clear directions are already visible this year—it is no longer a question of "whether VLA alone is enough," but how to make VLA work stably in real-world scenarios. The industry is no longer debating whether to use VLA or world models, but is shifting towards hybrid architectures; the focus of world models has shifted from "whether to use them" to "at which layer to implement them."
Therefore, our approach is very clear: VLA is the core skeleton, and the world model is the cerebral cortex that enables it to "think before acting." Before executing an action, the robot first internally simulates the physical outcome, predicting the trajectory of objects under force and the causal chain of multi-step operations, thereby avoiding failure at the source. This is not about replacing VLA with a world model, but using the world model to make VLA stronger.
Xinghaitu: VLA and World Models Share the Same Origin and Will Ultimately Converge The industry is currently somewhat overly focused on technical concepts, neglecting the technical essence. In our view, the core of both is interconnected. Whether it is VLA or a world model, the essence is transforming inputs into tokens, completing encoding through the multi-layer neural network of a Transformer, and then conducting pre-training via self-supervised methods. The basic logic is consistent.
The only difference between the two lies in the self-supervised training methods; world models primarily rely on video prediction for self-supervised training. They share the same origin and coexist, and will eventually converge in the future.
Lexiang Tech - Yuandian Robotics: A Single Model Cannot Solve the Real World; Robots Need a Continuously Evolving System This year, the industry is indeed forming a clearer consensus: a single technical route can hardly independently solve all the problems of embodied AI in the real physical world. VLA equips robots with the ability to directly map from vision and language to action, which is a very important foundation; however, when facing unstructured scenarios such as homes, relying solely on VLA is often insufficient. It also requires joint support from data closed-loops, task planning, motion control, anomaly recovery, and continuous learning mechanisms.
What aligns Yuandian's route with the industry consensus is the shared belief that robot intelligence must transition from "preset execution" to "data-driven" and "continuous evolution." The difference is that instead of merely stacking capabilities around a single model, we connect the data platform, training platform, Skill Hub, and the Bridge ontology through OpenBridge, an open-source developer ecosystem platform, enabling capabilities to be created, reused, shared, and iterated on the open-source platform.
II. Is the World Model the Next "VLA" Concept Stock? Everyone is talking about world models, but like VLA last year, it has not yet formed a stable definition. Some generate future videos, some predict the consequences of actions, some understand state changes in terms of events, and others simulate in latent space. People use the same term, but they are not necessarily building the same type of model. What is truly worth comparing now is what it actually inputs, what it predicts, and how it affects the robot's actions.
Unitree Robotics: The Biggest Bottleneck is the Deviation Between the Input and Output of AI Models and the Real World This year, the direction people have probably heard the most about is still the world model. In fact, we started working on video generation-based world models in early 2020, but by the end of 2020, the results were not particularly ideal, so we put it on hold for a while. Starting last year, we have vigorously resumed the development of video generation-based world models.
Currently, the biggest bottleneck in global embodied AI is that there is still a deviation between the input and output of AI models and the real world. For example, making a robot move, assembling things together, or moving an object to another position. From a macro trend perspective, current models actually do not have major problems and can roughly execute your tasks. The problems usually occur in the last few centimeters, or even the last few millimeters.
Zibianliang Robotics: VLA Answers "What to Do," World Models Answer "What Happens After" The world model receives data such as the robot's current and historical visual information, language tasks, actions, and proprioceptive states, predicting how objects and the environment will change after taking an action. VLA primarily answers "what to do next," while this capability helps the robot understand "what will happen after doing this."
The two are not in a substitutive relationship, nor is it just about one model scoring the actions generated by another; rather, they will gradually integrate within a unified model. Our WALL-WM aligns vision, language, and action through complete events, enabling the robot to simultaneously understand the action process, state changes, and final outcomes.
Keenon Robotics: Inferring Physical Outcomes in Latent Space Under the VLA Architecture With Keenon KOM 3.0, we integrated a latent space world model into the VLA architecture.
Service robots do not face fixed desktops in laboratories, but highly dynamic, unstructured environments like restaurants, hotels, and coffee shops. What robots need to do are often long-sequence tasks—for example, making a cup of coffee requires fetching the cup, receiving it, and delivering it; if any step goes wrong, the entire task fails. If relying solely on VLA for the end-to-end mapping of "perception-decision-execution," lacking the ability to predict physical outcomes, the success rate in complex scenarios will plummet.
III. Is Intelligence Proportional to Human Effort? If VLA is not the answer, and the world model is not a ready-made answer either, then the key question actually returns to the most fundamental place: data. Last year, the industry was still discussing whether the algorithms were inadequate or the data was insufficient. This year, the clearer answer is: the data is insufficient, which is why the algorithms are inadequate. Simulation, synthesis, and internet videos can serve as supplements, but for robots to understand the physical world, they ultimately must return to real interaction data.
Orbbec: Robot-Agnostic Data Collection Must Transition from Ad-hoc Solutions to Standardized Infrastructure Behind this wave of enthusiasm is first and foremost a real data shortfall. Embodied AI models need not just images and text from the internet, but massive amounts of real-world interaction data containing spatial information, actions, object state changes, and task processes.
At this stage, directly relying on the robot body for data collection usually faces issues such as high body cost, limited deployment efficiency, and hardware not being fully finalized. Robot-agnostic data collection can cover more scenarios, tasks and operational workflows at a lower cost. It is especially suited for capturing first-person environmental information, spatial relationships and long-horizon task sequences.
This type of demand comes from various types of customers, including embodied AI model companies, robot body manufacturers, professional data collection service providers, data operation companies, and other algorithm-driven enterprises. The common appeal of different customers is to gradually transform real-world data collection from ad-hoc equipment setups into standardized, deployable, and batch-deliverable infrastructure.
Genisom AI: Self-Developed and Self-Collected; External Data Cannot Replace Autonomous Operation Data from Our Own Robots We are working on human wearable collection, robot teleoperation, autonomous robot operation, and simulation data, because different data sources have their respective pros and cons.
Human wearable collection is mainly used for rapid learning; robot teleoperation is used to improve operational flexibility; because Genisom AI's main focus in the second half of the year is "getting the job done," autonomous robot operation is used to improve deployment outcomes and solve various failure issues encountered in the real physical world, which is an advantage that other data sources do not possess.
Simulation data is massive in volume and can help us expand the diversity of scenarios and tasks, but the sim-to-real problem prevents us from using this data directly; it all requires fine-tuning with real data.
For us, what we lack is not a specific type of data, but actually "useful data"; much of the massive amount of data available today is actually "invalid data."
Our data is all our own, including the data collection equipment, which is self-developed. In our view, data quality is one aspect; another is that the scenario distribution of externally provided data does not match—"their world ≠ our world." No matter how large the data scale, it cannot replace the interaction data generated by our own robots in real target scenarios.
Zibianliang Robotics: Robot-Agnostic Data Collection Expands Data Scale but Cannot Replace Real-Robot Data The current wave of enthusiasm is "robot-agnostic data collection," which means collecting human movements without relying on the robot body, using wearable devices such as headbands and gloves.
Robot-agnostic data collection is easier to scale than real-robot teleoperation and can cover more tasks and scenarios at a lower cost. The industry has basically solved the problem of "being able to collect," but from data replayability and transferability to truly improving model capabilities, it is still in the early stages. Zibianliang Robotics has launched the QUANXTA Zero robot-agnostic data collection solution, which features full-body mobile data collection and mobile collection data body replay capabilities, with a data ingestion efficiency above 85%.
However, it cannot completely replace real-robot data because human body structure and movement patterns differ from those of robots, making it difficult to fully present the contact, errors, and failures during robot execution.
From a technical route perspective, world models can also reduce reliance on real-robot demonstration data for single tasks. VLA requires more demonstration data with accurate alignment of language instructions, robot observations, and action trajectories; world models, on the other hand, focus more on the continuous states and event evolution before and after an action occurs, and can use data from different sources such as real robots, robot-agnostic, videos, as well as simulation and synthesis. What it truly solves is not just reducing data volume, but improving data utilization efficiency, enabling the model to learn transferable physical laws, thereby enhancing cross-scenario generalization and long-horizon task capabilities.
Keenon Robotics: What is Collected is Just "Crude Oil"; the Real Bottleneck is "Data Refining" Human data cannot completely replace real-robot data; the relationship between the two is complementary rather than substitutive.
Human first-person perspective data can efficiently collect the complete process of humans completing tasks, used to train foundation models and build scenario cognition and task understanding capabilities. But ultimately, to enable the robot body to execute precisely, real-robot data is still needed for post-training fine-tuning—because real-robot data contains signals that are completely absent in human data, such as the kinematic characteristics, joint torques, and force feedback of the specific robot body.
Collecting data itself is not difficult now; the difficulty lies in making the collected data truly improve model capabilities. It's not as simple as wearing a headband, recording a few hours of video, and using the data directly. From raw video to trainable structured data, it must go through a series of processing steps such as undistortion, pose estimation, depth reconstruction, hand-object segmentation, and automatic annotation. If this entire "data refining" capability cannot keep up, collecting more raw data is just collecting "crude oil," which cannot be turned into the "fuel" to drive the model.
The "sim-to-real" gap from simulation to reality is particularly evident in service scenarios—no matter how realistic the physics engine is, it cannot simulate the subtle differences in the water outlet of a real coffee machine, the changes in friction on the surface of cups made of different materials, or the "accidents" in the real world such as a customer suddenly reaching out to interfere with the operation trajectory.
IV. Can Perception Data Fill the Gap? What was first discussed at scale this year is still visual data: first-person perspective, RGB-D, wrist near-field, multi-view, IMU, and time synchronization. But at the same time, data closer to physical interaction, such as tactile, force, slip, material, and contact feedback, has also begun to emerge from the supplier side.
YIMU Intelligence: Visuotactile Sensing, Turning Every Contact into "Physical Ground Truth" The visuotactile route is the most human-like route. It is not a traditional force sensor, but establishes an optical observation system inside flexible materials. When the robot contacts an object, it uses images to capture the micro-deformations of the material, and then converts these deformations into physical signals such as pressure, friction, slip, hardness/softness, and texture.
Physical AI cannot rely solely on making models continuously larger; what truly matters is whether robots can continuously obtain physical ground truth from the real world. Every contact is a learning experience for the robot to understand the world; every grasp generates new Ground Truth.
If a world model is trained purely on video, it can learn visual variations yet fail to capture what transpires at the moment of physical contact. The significance of tactile data is to turn the physical results when the robot contacts an object into something the model can learn. TouchNet, which YIMU Intelligence is developing, aims to turn data such as pressure, texture, sliding, material, and force feedback in real contacts into reusable data infrastructure.
PaXini: Tactile Sensing is Both a Robot's Capability and Data for Model Training Perception is the first gateway for robots to enter the physical world. Starting from tactile sensors, PaXini's business continues to extend to dexterous hands, robot bodies, datasets, and models, hoping to connect perception, decision-making, and execution into a complete technical ecosystem.
PaXini is working on two fronts simultaneously: tactile sensors and dexterous hands on the robot side, and a human wearable data collection system, allowing humans to wear tactile collection gloves to complete tasks, synchronously recording multimodal information such as tactile, visual, and action trajectories to cover real scenarios with higher efficiency.
For example, picking up a cup, screwing in a screw, and organizing clothes all rely on stable, controllable contact. In these fine operations, tactile sensing has become a fundamental capability: it not only affects whether the action succeeds but also records the physical results generated by the contact for the model.
Orbbec: Hardware Determines the Starting Point and Upper Limit of Data Quality This year, we launched EGO, UMI, WristCam, and Hub, hoping to cover key perspectives in real interactions. EGO mainly solves first-person perspective and environmental spatial information collection; WristCam and UMI are closer to the hand and operation areas, used to supplement wrist near-field observations and hand-object interaction details under occlusion conditions; Hub is responsible for multi-device synchronization. Customers often do not choose a single device independently, but combine different devices into a collection solution based on tasks, perspectives, and data modalities.
From the perspective of customer value, what customers ultimately need is certainly not the equipment itself, but data that can effectively improve model capabilities. However, "trainable data" does not begin to form only after collection is completed. Sensor configuration, perspective design, depth quality, time synchronization, and calibration accuracy will all directly determine whether subsequent data can be used and how much cleaning and correction cost is required.
Therefore, while customers superficially purchase hardware and customized services, what they actually pay for is the capability to continuously acquire high-quality, reusable training data. Hardware is not the ultimate goal, but it determines the starting point and upper limit of data quality.
Currently, the industry has basically crossed the first stage of "whether data can be collected" and is entering the stage of "what data to collect, how to evaluate data, and which data truly improves models." The focus of the next step in competition will not just be the number of devices or the total amount of data, but whether the data possesses accurate spatial information, complete task chains, stable multimodal synchronization, and whether its effectiveness can be proven through training results.
If future world models are to truly understand the physical world, tactile and force sensing are unlikely to remain just marginal signals in control systems. They will become part of the world models.
At this year's crowded WRC, you actually feel that embodied AI companies have become quieter, and the arguments about algorithms have given way to the same thing: collecting data.
Over the past two years, the industry has been accustomed to understanding embodied AI as putting a body on a foundation model: cameras are responsible for seeing, joints for moving, dexterous hands for executing, and everyone's attention is focused on the "brain."
When robots actually start entering restaurants, warehouses, and homes, the body, in turn, bottlenecks the brain. The deviations in the last few centimeters, the slip at the moment of contact, and the friction of different cups cannot be directly grown from language and internet videos. The robot body once functioned merely as the model’s output interface; it is now emerging as an input source for intelligence.
Therefore, the real change in the exhibition halls this year happened in the places that look least like the "brain": headbands, wrist cameras, tactile gloves, dexterous hands, teleoperation consoles, and data factories. Before finding the ultimate route, they are first feeding the physical world to the models piece by piece.