In the first half of 2026, the Embodied AI sector attracted over CNY 46 billion in funding. Humanoid robot manufacturers also pulled out all the stops, with a new robot demo going viral across the internet every few weeks.
However, behind the hustle and bustle, the entire industry has hit an invisible wall: the real-world data collected from physical robots is too expensive and too slow to feed the increasingly hungry embodied models.
At the recently concluded "Hardcore Launch — Hard Tech Investment and Financing Development Forum," He Yonghan, CTO of Intime Technology, delivered a speech titled "Empowered by Simulation: Bridging the Last Mile for Embodied AI Deployment," providing Intime Technology's answer.
Key Takeaways (30-Second Overview)
- Simulation is not a cheap substitute for physical robot data, but a data format that physical robots can never obtain—it inherently possesses "privileged information."
- Six-dimensional comparison: kinematic fidelity, sensor configuration, use cases, production cost, storage cost, and paradigm risk show simulation holds a comprehensive advantage over Ego data.
- The essence of the Sim2Real Gap is parameter deviation, not rule misalignment: humans don't need to relearn how to walk on the Moon.
- IntimeVerse creates a true closed loop: Code2Space → Asset2Physics → Space2Data → Training → Evaluation → Edge Deployment → Physical Robot Alignment → Upstream Feedback.
- Already doing real work: parcel induction, power battery assembly, factory inspection, daily life supermarkets... Projects have entered the order delivery stage.
Reading the Table: Embodiment, Models, and Tasks
He Yonghan set the tone for the sharing right at the beginning: "This is a story about embodied data."
Regarding the embodied models that the industry cares about most, he highlighted two of the hottest terms: VLA (Vision-Language-Action model) and VLN (Vision-Language-Navigation model). One is for manipulation and the other for navigation, both stemming from the academic ambition to "solve complex manipulation problems with a single end-to-end model policy."
"However, as the field has evolved to its current state, people are increasingly finding that if they rely solely on visual observation data, many tasks simply cannot be accomplished."
Language is more of a source of intent rather than the basis for manipulation. The approach of world models is to equip the model with the ability to predict future world states, and then feed the predicted features back into current decision-making.
He Yonghan's judgment is that these two lines are currently basically converging.
As for tasks, from a technical perspective, there are only three major categories: Mobile Manipulation, Navigation, and pure Manipulation.
There is a sentence that is actually quite heartbreaking: most of the silky smooth demos you see at WAIC and WRC only cover one-third of them.
The Data Pyramid: The Academic One Should Be Torn Down
He Yonghan showed a diagram from a review—currently, the sources of embodied data are universally recognized in the industry as a five-layer pyramid, but Intime Technology has compressed this pyramid into three layers.
The bottom and largest layer is internet data, which is extremely low in cost but not high enough in quality, mainly used to train multimodal foundation models. Above it is simulation data, the volume of which is already second only to general data. Further up is the Ego data (egocentric perspective data) that has been particularly popular in the past two years: detached from the robot embodiment, humans wear headbands with multiple cameras to record the entire workflow.
Further up is UMI data: the headband is still worn, but the end effector is replaced with a device isomorphic to the real robotic arm, so the collected data has no configuration gap with the physical robot. In contrast, Ego records the dexterous manipulation of human hands, which has a huge configuration difference with the end of the robotic arm.
At the top of the pyramid, with the highest value, is teleoperation data: through isomorphic teleoperation or VR teleoperation, the robot is directly controlled to complete tasks in real scenarios, and the robot's own data is recorded.
Different companies have different understandings of this pyramid, which has evolved into different data solutions. And He Yonghan gave a very bold judgment:
"With simulation's evolution to date, it already possesses the potential to replace all the bottom layers in the aforementioned pyramid."
Intime Technology's data pyramid has only three layers:
Layer 1: Massive simulation data. No restrictions, freely collected across scenarios, embodiments, and tasks. The largest volume, which can be used for pre-training.
Layer 2: Scenario-specific simulation data. Data is "narrowed down" to serve a specific scenario, but can still be collected across embodiments, used for post-training.
Layer 3: A small amount of real-world data. When it comes to actual deployment, collect a little real-world data for a small amount of teleoperation data, used for post-training to align with the real world.
Taking Ego Data as the Opponent, Where Does Simulation Win?
Kinematic fidelity. Ego data must undergo post-processing: estimating hand poses from videos. No matter how accurate this algorithm is, there is deviation—recovering 3D information from 2D images, the error is physically unavoidable. Then the estimated pose needs to be aligned with the target embodiment's end effector, adding another layer of error.
In simulation, this is hardly a problem: hardware manufacturers provide embodiment description files for every piece of hardware they release, 1:1 with the real embodiment, with degrees of freedom and dynamic parameters for each degree of freedom truly aligned.
Sensor fidelity. Ego basically only has vision, lacking force and tactile senses, and the camera's intrinsic and extrinsic parameters cannot align with the physical robot. Simulation is completely flexible: vision, force, touch, add whatever you want.
Data usage. Because of the first two drawbacks, Ego data is currently mainly used for pre-training; simulation data can be used for both pre-training and directly entering post-training.
Production cost. The industry's cost of producing 1 hour of physical robot data is about CNY 200 to CNY 400. Simulation is currently below CNY 200, and with the accumulation of scenarios and assets, the cost will drop rapidly. The generation cost for a single scenario has dropped from hundreds of US dollars to under US$10, and the cycle has been compressed from weeks to minutes.
Storage cost. This cost, which is "very likely to be ignored by the vast majority of people," involves Ego collecting videos, where a simple task can easily take up hundreds of gigabytes. According to market rates, the cost of 1PB storage for a year is CNY 600,000 to CNY 800,000, and it's not even the highest-speed storage. Calculated at a future scale of 100 to 1,000 PB, the bill would be outrageously high.
Simulation's solution is to trade compute for storage: instead of storing data in advance, it runs many simulation scenarios on servers and collects data directly within the scenarios—spending money on computing power, which is much cheaper than storage.
Paradigm risk. This is "a very profound lesson" from He Yonghan's many years of working on models: when the model paradigm has not converged, it can change at any time. Once it changes, there is a high probability that the requirements for data will undergo fundamental changes—observation frequency increased from 10Hz to 60Hz, resolution pulled higher, action representation changed.
Data collected in advance in the real world, when faced with such changes, is very likely to quickly drop to zero value.
But simulation won't. When conditions change, you can adjust them in the simulation at any time: perspective, resolution, frame rate... everything can be adjusted.
An industry actual measurement reference is: in complex manipulation tasks, training with simulation data 5 times the scale of physical robot data can increase the robot's real-world manipulation success rate by 40%, while the acquisition cost is only one-third of that for physical robot data.
Sim2Real Gap: Do Humans Need to Relearn Walking on the Moon?
After talking so much about the benefits of simulation, there is a problem that cannot be bypassed.
What comes out of simulation will always have differences from the real world. This is a 100% existing problem, and it cannot be avoided no matter how good the simulation is.
But He Yonghan believes this problem needs to be viewed from a different perspective, and the answer has two layers.
First layer: What exactly are we learning in simulation?
The answer is—the framework.
He borrowed the analogy most loved by CEO Wang Deshui: the gravity on Earth and the Moon is different, but they follow the same set of physical laws.
Earth can be analogized to the real environment, and the Moon to the simulation environment. Regarding "how humans walk on the Moon," the walking posture is the same. Even if arriving at a completely new environment like the Moon, only a slight adjustment to adapt to the physical parameters is needed, and the problem is solved.
What the model needs to learn is the framework of how the physical world operates, not the precise value of every parameter.
Second layer: The Gap itself can be compensated by algorithms.
After the model is trained in simulation and brought to the real world, a small amount of physical robot data can be used for fine-tuning and alignment, or RSI (Recursive Self-Improvement) can be used for rapid calibration.
He Yonghan's judgment is: on the premise of a sufficiently large amount of data, within the next 1 to 2 years, the industry will usher in an inflection point of intelligent emergence driven by simulation data.
IntimeVerse: The Closed Loop for Embodied Deployment
For methodology to land, it relies on a system. Intime Technology's answer is called IntimeVerse, a complete chain:
Scenario Generation → Asset Generation → Data Collection → Model Training → Simulation Evaluation → Physical Robot Deployment and Alignment → Problem Feedback to Upstream for Modification.
Breaking it down into four modules.
Code2Space: Letting the model write "construction code" instead of drawing renderings
Both inputs work: if there are images and videos of real scenarios, restore them into the simulation as much as possible; if there is nothing at all, just giving a text description can also generate a realistic simulation environment.
The key lies in the working method—not directly reconstructing the environment, but using a foundation model to write a piece of code that can generate this environment, and then the simulator executes the code to "build" the environment.
Intime Technology trained a vertical model specifically for this on top of the foundation model, with the diversity dimension pulled very wide: lighting, object layout, object types, everything can be changed.
Especially for desktop scenarios, a separate direction was made, because a large number of fine manipulations occur on desktops.
Our approach has a huge advantage in time and cost over traditional manual scene creation by artists.
In May this year, Intime Technology jointly launched AnySceneGen with D-Robotics, promoting the scaled generation of simulation scenarios and assets.
Asset2Physics: Turning "looks like" into "feels right"
The biggest difference between embodied manipulation and autonomous driving is: autonomous driving tries its best not to touch things, while embodied manipulation is exactly the opposite—it is precisely about touching and interacting. Once contact is made, the physical properties of the object must be accurate: volume, collision geometry, mass, center of mass, surface friction coefficient, every single one will affect its reaction in the physics engine.
Intime's solution is a set of differentiable parameter identification: after the object is generated, the robotic arm first performs operations like pushing, pressing, and flipping on it, observing whether the state changes meet expectations; if not, it automatically analyzes which parameters need to be adjusted, and goes back to modify them, iterating like this until the difference from reality is small enough.
It's not about labeling 3D objects, but using interaction to compile physical objects into simulatable matter.
Currently, single-asset physical calibration can be done at the minute level, with appearance and physical accuracy exceeding 95%.
Space2Data: The biggest bargain of simulation is that it "knows everything"
With the environment ready and physics correct, the most difficult bottleneck arrives: who is going to do the work in this scenario and collect the data?
Intime Technology has two paths: autonomous task discovery (letting AI discover possible key tasks in the scenario on its own, then organizing execution and recording trajectories) and executing according to SOP (in manufacturing, tasks are often pre-defined, and going through the SOP once can also collect data).
Some people might ask: since you can collect autonomously in simulation, why not just throw the robot into the real environment to collect on its own?
The difference lies in—simulation possesses Privileged Information.
The agent always knows where everything in the scenario is, and knows it at the coordinate level, which you cannot do in a real scenario.
Intime precisely uses this information gap to obtain high-quality data, and then reversely trains robots that can operate autonomously in the real world.
This system currently has a task execution success rate of over 90% in simulation, and a single RTX 4090 graphics card can produce over 5 hours of data per day. Both successful and failed trajectories are retained, and failed samples can also locate insufficiently covered environmental conditions and action links, guiding the next round of collection—the data gets better and better with use.
Data2Edge: Instead of making the brain bigger, make the cerebellum faster
After the model is trained, it is first evaluated in simulation—evaluation in the real world is too costly. If it is "very poor" in simulation, continue iterating; if it's good enough, then enter the real environment. This is also the most common thing currently done in simulation in the industry.
When it comes to deployment, He Yonghan gave a sober judgment:
On the brain side, it is very likely to be consumed by foundation models in the future. So we pay more attention to edge-deployable, real-time, cerebellum-leaning models.
Tasks like parcel sorting do not require language at all; they just execute high-frequency repetitions. The essence is the logic of the "embodied cerebellum."
Intime Technology's self-developed IntimeBrain, through model structure innovation, achieves 20Hz operation on edge chips with 100 TOPS computing power—mainstream solutions either cannot run on the edge or only have 1Hz, which is one decision per second. Later, they plan to increase the frequency to 100Hz, which is basically on par with the control frequency of the robotic arm itself.
Real Work: Scenarios Already Running
Facing customers, the entire closed loop is packaged into two major stages: Real2Sim (moving the customer's specific scenario and robot embodiment assets into the simulation and restoring them) and Sim2Data (making the robot move as expected in the scenario and collecting the data). The collected data includes multi-view vision (head + dual wrists), embodiment degrees of freedom (joint angles of each joint), as well as force and tactile sensor data—as long as they are installed, everything is recorded.
Parcel induction. The hottest scenario in the embodied track this year, with a bunch of top giants doing POCs in it. The difficulty for simulation lies in: the environmental states of parcel induction vary greatly among different companies, and the parcels are also complex—paper boxes are closer to rigid bodies, packages are closer to soft bodies, and there are also irregularly shaped parts, with significant differences in mass and friction characteristics. Moreover, the parcels are continuously moving and colliding with each other on the conveyor belt. Around this scenario, Intime Technology has created the industry's first 100% simulation-based full-process training solution for parcel sorting.
Power battery assembly. A high-value vertical category considered "very promising in the future" after communicating with customers. The things operated on here are closer to rigid bodies, and the physical properties are relatively simple. The only difficulty lies in one point: the screwing in and unscrewing of screws in holes, whether this contact process can be simulated accurately. Currently, Intime Technology is providing simulation training data and related services according to the production line needs of a certain new energy battery assembly enterprise.
Factory inspection navigation. Give the robot dog a semantic instruction, and it navigates autonomously in the factory, taking photos or performing sensor detection upon reaching the waypoint. This type of task covers all three categories: mobile manipulation, navigation, and manipulation.
At the same time, Intime Technology is also advancing simulation training data services for the specific robot configuration and manipulation tasks of a top embodied AI manufacturer in Beijing, and has entered large commercial supermarket scenarios. Related projects have entered the order delivery stage.
Conclusion: How Long is the Last Mile Really
What this speech really wants to sell is not "cheaper data," but time.
He Yonghan repeatedly emphasized that the model paradigm has not yet converged, and data requirements could be overturned at any time. In other words, physical robot collection buys "past requirements," while simulation buys "future possibilities."
In a track where the paradigm changes every month, the ability to start over is more valuable than getting it right the first time.
More notably, his revelation of "privileged information" is crucial. Many people's attention is focused on whether the visuals are realistic enough, but the Ground Truth is the watershed—collecting data in the real world requires paying the triple price of annotation, pose estimation, and alignment, while simulation inherently possesses it for free. This is not "cheaper physical robot data," but a data format that physical robots can never obtain.
An open question is: will embodied AI follow the old path of autonomous driving—first a carnival, then discovering the long tail, and finally slowly grinding it out through simulation and closed loops? If so, then whoever holds the "low-cost, reproducible, closed-loop" simulation training system today holds the ticket to the next round of competition.
Intime Technology is betting on this very ticket.
About Intime Technology
Intime Technology is a startup focusing on embodied AI simulation data and training systems. It has consecutively completed its angel round and angel+ round of financing within two months, with a cumulative financing of tens of millions of CNY.
The core team is a lineup where academia, engineering, and deployment are closely complementary:
Founder and CEO Wang Deshui: Former head of large model technology at Amazon Alexa, leading the team to advance the Code-to-Space technical route, building editable, interactive 3D assets and simulation spaces with physical properties.
CTO He Yonghan: Ph.D. from the Institute of Automation, Chinese Academy of Sciences. Former head of embodied AI algorithms at D-Robotics. Responsible for the full-link R&D from simulation data generation, VLA/VA model design to physical robot data infrastructure construction and edge deployment. Published over 30 papers in international top conferences and journals such as ECCV, CVPR, ICRA, and IROS, and holds over 10 authorized patents.
Chief Scientist Su Hu: Associate professor at the Institute of Automation, Chinese Academy of Sciences. Long engaged in theoretical research on control science, computer vision, embodied AI, and machine learning. His related work is one of the earliest representative achievements in generating interactive simulation scenes and synthesizing robot manipulation data through text or a single image.
In terms of commercial landing, Intime Technology has chosen an extremely pragmatic route: prioritizing industrial manufacturing scenarios. The reason is exactly as Wang Deshui said—"Physical AI is directly related to the physical world, and the consequences of problems are much greater and more intuitive than those of language models." Industrial scenarios are relatively fixed and have a high foundation of automation, making them more suitable for the testing and verification of new technologies.
The value of embodied capabilities lies not in whether the form looks like a human, but in whether it can handle complex situations and complete tasks stably.
From AnySceneGen to IntimeVerse, from spatial generation to customer delivery, Intime Technology is turning the judgment that "simulation is the only solution" into products and orders.