谢晨 on Robotics’ Data Drought, Synthetic Data and Meta’s Scale Deal
Summary
谢晨 reads Meta’s roughly $15B payment for Scale AI as proof that “the ticket to entry for a future $10T company is mastering AI data.” In his view, the transaction looks more like an acqui-hire at a roughly $30B valuation: Meta is buying Alexandr Wang’s understanding of data, talent network and organizational ability. His prediction that Wang will “most likely succeed Zuckerberg in 10 years” is extremely aggressive, based on the idea that Wang may be giving up Scale’s potential 10x upside over 3–5 years to pursue greater personal impact.
The data economics of embodied intelligence are the inverse of autonomous driving: the former must rely mainly on synthetic data, while the latter has mass-produced fleets continuously collecting real-world data. Cruise used synthetic data for roughly 30% of its overall volume; for an individual long-tail scenario, each case might require 1,000–10,000 synthetic samples. 谢晨 also says synthetic data now accounts for more than the 1:99 or 1:100 ratio cited for ordinary scenarios. Robots have no million-unit installed base, and if the industry relies mainly on teleoperation, the business model must first close a positive loop. His conclusion: the “synthetic-data-first” route is more likely to work.
The core metric for simulation is not visual fidelity, but whether adding data improves the model and whether the training result truly transfers sim-to-real. At Cruise, 谢晨 built an evaluation system first, then spent roughly six months locating and fixing issues in distributions, sensor noise and other factors to produce an effective solution. He compares high-utility data to “strong coffee,” arguing that less data should still deliver more information gain. Much of today’s simulation is more like “water that gives you diarrhea”: it can be generated, but it cannot train or deploy a model.
Embodied simulation is harder than autonomous-driving simulation, with bottlenecks in cross-embodiment, physical interaction and vision-in-the-loop RL. A car’s interaction with the ground is relatively simple; a robot must handle opening refrigerators, picking up straws and typing on keyboards. A real hand may lift 1–2 kg, while an incorrect simulation model may manage only 0.1 kg. Closed-loop RL also requires many perceptual environments to run in parallel on a single GPU, so accuracy, total FPS and the feedback loop must all work together.
谢晨 expects the embodied industry to split into four layers: hardware, foundation models, vertically integrated software-and-hardware companies, and simulation-first infrastructure. The foundation-model layer may ultimately require more than 100,000 GPUs, massive datasets and the world’s top talent, giving large technology companies the advantage. Vertical companies such as Figure will capture value through fine-tuning and deployment. US-China differences in willingness to pay will further shape company boundaries: the US favors specialization, while China’s weaker software monetization may push companies toward “big and complete” businesses that monetize through hardware.
The real-data camp and simulation camp, despite their apparent differences, are converging on a Data Pyramid: internet data at the base, synthetic data in the middle and a small amount of real data at the top. 谢晨 says co-training on real and synthetic data is usually better than relying on either route alone; the remaining differences are mostly in the mix, architecture and RL weighting. “Good data is the product of two things”: diverse scenarios multiplied by high-quality human demonstrations. The best teacher ultimately does not repeat demonstrations, but “sets questions” for the model to solve through RL.
Nvidia treats simulation as a third computer alongside the data center and edge physical-AI systems, leading 谢晨 to estimate a potential market in the “trillion-to-$10T range.” “NVIDIA is a simulation company” extends the company’s underlying logic from game graphics to physical AI. Isaac Sim and MuJoCo may also gradually converge through MuJoCo Warp and Newton. The end state is not merely cross-embodiment, but “cross-universe”: game data could become pretraining material for models that generalize across worlds.
Embodied intelligence is still at the “GPT-1 stage,” with no stable Scaling Law yet, but 谢晨 is betting that a breakthrough could appear within 1–2 years. His rationale is that talent density, capital attention, and the lessons from GPT on Transformer architectures and scaling data and compute are all much stronger than in the early autonomous-driving era. The real risk is not being early; it is allowing financing and demos to become a team’s “reward model,” with valuation replacing customer revenue and technical progress.
Deep dive
1. The Data Drought Is Turning Simulation into Embodied AI Infrastructure
谢晨 defines sim-to-real as the unavoidable gap encountered when a robot algorithm trained in simulation is deployed on a real body and in real-world environments. Synthetic data is generated programmatically through simulation, algorithms or a combination of both.
Guanglun Intelligence is taking a hybrid approach, focusing on 3D data, robot trajectories and even human-in-the-loop data rather than just the 2D labeled images common in early computer vision.
Founded in 2023, the company is currently at Series A with an undisclosed valuation. 谢晨 says it is still loss-making but “extremely close to profitability.” The episode frames data as the key bottleneck in embodied-AI training and scaling, and 谢晨 expresses strong conviction in synthetic data.
2. Cruise’s Matrix Proved That Looking Real Is Not the Same as Being Trainable
谢晨 joined Cruise in 2018, initially working on high-definition maps. He says he rebuilt the update, production, augmentation and quality-control pipelines in roughly 3 months, after which he was moved to take over the simulation group, a function under direct pressure from the CEO.
Cruise initially hired game engineers to recreate San Francisco using technologies such as Unreal and named the system Matrix. The problem was that the data could not be handed directly to the perception team. It was “more like a great-looking toy.”
The fundamental mismatch was the objective function: game engineers optimized for spectacle and excitement, while machine-learning teams needed to know whether feeding generated data to a model actually produced a meaningful improvement.
3. The Breakthrough Took Six Months and Started with Evaluation, Not Generation
After taking over, 谢晨’s first move was not to improve image quality or generate more data, but to build an evaluation system. With measurable standards, the team could determine whether the sim-to-real gap came from visuals, distributions, sensors or noise.
The first category of standards measured absolute realism: synthetic data was benchmarked against real data for lighting, color, sensor noise and label accuracy.
The second measured utility: could the data improve the model after entering training? “You can buy a cup of strong coffee, or a cup of iced coffee diluted with a lot of water.” Future data competition will focus more on information density per sample.
The team also used an early version of GANs to transfer synthetic data into a more realistic style, such as making the sky grayer and sensor noise heavier. After roughly 6 months, it had an effective synthetic-data solution that improved Cruise’s perception models.
4. A Usable Data Pipeline Must Close the Loop with Quality Control
Pipeline inputs include scenarios defined by humans or observed in the real world, which are then generalized through statistical, manual and AI-based methods. Static assets such as bicycles, cars, roads and trees are combined with dynamic behavior to form the world.
Sensor simulation converts that world into data observed from the vehicle’s perspective, after which the system automatically generates 2D, 3D and even 4D labels, along with object semantics and other metadata.
谢晨 emphasizes the final quality-control stage. Only when automated checks and human-in-the-loop checks jointly validate the output can the data reliably support training; completion of production is not proof that the data is valid.
5. Long-Tail Scenarios Turn One Real Accident into 10,000 Experiments
A homeless person suddenly darting into the road in San Francisco, steep slopes and heavy fog are all critical scenarios, yet rare in the real-world distribution. Without samples, teams can neither train effectively nor uncover system-level safety problems.
For one pedestrian intrusion, simulation can vary age, clothing, posture, lane, traffic, weather and lighting, generalizing a single sample into thousands of failure modes and running the algorithm through them 10,000 times.
Once a failure is identified, precise labels feed the case back into the perception or prediction model, creating an iterative loop: discover the problem, expand the problem, fix the model and validate again.
6. Autonomous Driving’s 30% Synthetic Share Is Reversed in Embodied AI
Cruise used synthetic data for roughly 30% of its overall volume. But for a corner case encountered only a few times a year, 谢晨 believes the ideal ratio is 1,000–10,000 synthetic samples per long-tail scenario.
For ordinary scenarios, he says the synthetic-data share exceeds the 1:99 or 1:100 level cited by 王鹤. That statement refers to a different ratio from the per-scenario allocation for long-tail cases.
Autonomous driving has a century-old vehicle platform. Mass-produced fleets from Tesla, Li Auto, Xiaomi and others continuously return data; owners do the driving, while automakers mainly pay for compute and bandwidth. The L2+ market will therefore remain primarily real-data-driven.
Embodied robots have no installed base of millions or tens of millions of bodies operating freely in the world. 谢晨 concludes that the industry must initially rely on large volumes of synthetic data and small amounts of real data. “Otherwise, the business model simply cannot work.”
7. Teleoperation Must Clear Three Hurdles: ROI, Geography and Organization
张小珺 asks whether industry success inherently requires large volumes of real data, with synthetic data being rationalized only after the fact. 谢晨 concedes that teleoperation may work in small settings, but only if the value created by the robot closes a positive commercial loop.
Domestic automakers are already highly automated. Overseas factories may face restrictions on cross-state teleoperation, unions and internal politics. An outside robotics company deploying teleoperation for an automaker does not automatically create a positive business loop.
The value of teleoperation in hotels and retail may also be insufficient relative to cost. 谢晨 gives Tesla the strongest benefit of the doubt because it might optimize the process inside its own factories. He considers it “unlikely” that the entire industry can be primarily real-teleoperation-driven and still outperform synthetic data.
8. Nvidia Showed 谢晨 the Opportunity in China’s Mass-Market Autonomous Driving
谢晨 moved from Cruise to Nvidia to expand beyond a single L4 company and serve mass-market L2 customers. He also already viewed Nvidia as “the strongest hard-tech company in the world,” not merely a gaming-GPU vendor.
Before joining, he expected Waymo and Cruise to be Orin’s biggest customers. Once inside, he saw that the top 3 were “蔚小理” and realized that hundreds or thousands of L4 vehicles could not match China’s mass-produced fleet. The episode also notes that a vehicle at the time might carry 4 Orin chips, while Li Auto used 2.
German and Japanese automakers were more conservative on high-compute chips. Tesla developed its own in-car chips, although it still buys Nvidia GPUs in large quantities for data centers. 谢晨 therefore sees China as “far ahead” in L2+, while acknowledging Tesla as the strongest exception.
9. L4 and L2 Simulation Serve Reinforcement Learning and Imitation Learning
L4 must reach or exceed human performance in extreme situations, which made reinforcement learning important relatively early. It therefore emphasizes closed-loop simulation, agents and loop efficiency. L2 mainly seeks human-like driving, with imitation learning as the underlying logic and large-scale replay of real logs for validation.
Embodied intelligence will also move gradually from imitation learning toward more RL, but its RL and physical-interaction requirements are far higher than autonomous driving’s. The vehicle-side pipeline cannot simply be copied.
谢晨 also believes autonomous-driving teams may have over-expanded. Using Tesla’s roughly 240-person team as an example, he argues that excessive scale can leave insufficient room for breakthroughs in core foundational algorithms, trapping the organization in module-level optimization.
10. Nio Was a Landing Point Before Entrepreneurship; Robotics Was Always the Destination
谢晨 worked at Nio for roughly 1.5 years. When he joined, he had already told 李斌 that he ultimately planned to start a company. The expectation was to solve the immediate problems first and complete the handoff later.
After roughly 1 year, the team was using large-scale simulation evaluation to support second-generation models such as the ET7. Over the following 6 months, he completed the transition and invited a former Baidu simulation lead to take over.
His earliest startup deck was titled “Simulation for Robotics,” before “embodied intelligence” became a popular term. Investors at the time would ask whether robotics meant robot vacuum cleaners, factory arms or humanoid robots, which few companies were then building.
11. Autonomous Driving Gave the Company Its Quality System for the Embodied Transition
Guanglun first served mass-market autonomous driving, using synthetic data to fill edge-case requirements. 谢晨 lists NVIDIA, Wayve, Changan, Geely and BYD among its customers.
As embodied AI moved from storytelling into serious development, Guanglun’s first overseas embodied customer was Figure, followed by Nvidia’s Isaac team, Google DeepMind, international universities, and Chinese groups including BAAI and Galbot.
谢晨 views autonomous driving as the team’s training ground. It built a culture of scaled production, quality control and high-pressure delivery, but success on those “lower-level problems” does not directly solve the higher-level problems of embodied AI.
12. Cross-Embodiment Is the First Wall in Embodied Simulation
Autonomous-driving bodies differ relatively little: vehicles generally have 4 wheels, and changes in vehicle type do not rewrite the interaction paradigm. An embodied model ideally needs to work across different humanoid robots, robot arms and other bodies.
The same task must therefore handle different joint structures, dynamics, hand shapes and control interfaces. 谢晨 summarizes the objective as cross-embodiment, with simulation responsible for producing accurate data across those differences.
The second wall is physical interaction. Cars mainly contact the ground and try to avoid collisions; robots must actively pick up coffee, retrieve straws, open refrigerators and type on keyboards. Their interactions are “infinitely varied.”
13. A Single Refrigerator Door Exposes the Difficulty of Physical Real-to-Sim
Real-to-sim is not merely reconstructing the appearance of a real object as a 3D model. It maps real scenes, assets and physical properties into simulation, assuming that the physics engine is good enough and every asset carries the correct parameters.
A simulated refrigerator may only look like a refrigerator and may not even open. Even if it does, the hinge, mechanics, weight and force required to open the door may all be inaccurate.
Even with those details, the simulation may contain only one refrigerator model rather than enough assets matching the real-world distribution. Teams must therefore collect real physical information and put it into simulation rather than relying on visual reconstruction alone.
14. Vision-in-the-Loop RL Ties Accuracy Directly to Throughput
Imitation learning only requires collecting open-loop trajectories. If a person opens the cap of a tall bottle, a robot can imitate the motion; change the bottle to a short one or alter the rotation, and generalization may fall off quickly.
Early robot RL focused on locomotion. Today it must cover the entire body, especially the hands, and let robots observe and experiment in vision-in-the-loop environments.
谢晨 compares the path to large models: first build a foundation model, then fine-tune it with RL. Simulation must be closed-loop and massively parallel, with each GPU supporting enough perceptual environments. Otherwise total FPS will be insufficient for training.
15. Good Simulation Should Be as Drinkable as Water from a Tap
The first standard for successful simulation is sim-to-real: an algorithm trained in simulation must transfer to a real robot. Assets, scenes, solvers, APIs, metadata and the training toolchain must all be correct at the same time.
In real-world simulation, refrigerators often cannot even open. Even when they can, the hinges, weight, pulling force or asset distribution may be wrong. One visually beautiful asset is still insufficient to train a model that generalizes.
The robot model itself can also be distorted. A real robot or hand may lift 1–2 kg, but in simulation it may manage only 0.1 kg. One customer spent more than 6 months debugging without achieving alignment.
谢晨’s vision is to make high-quality simulation as universally available as drinking water. Today’s industry can produce water at scale, but “the water actually gives you diarrhea.”
16. Guanglun’s Moat Combines Physical Reconstruction, Solvers and the Deployment Loop
In real-to-sim, Guanglun uses AI to process information returned by physical-data-collection equipment and robot interactions, automatically estimate core physical parameters, and generalize them across large numbers of assets and scenes according to their distributions.
谢晨 claims that the foundational simulation materials used by the world’s best institutions are all produced by Guanglun, and that Guanglun is their “only simulation supplier.” This is the guest’s direct characterization of the company’s competitive position.
At the simulation layer, Guanglun uses ecosystems including MuJoCo and Isaac Sim while developing its own underlying solver on Warp. Google and NVIDIA are both customers and partners.
At the sim-to-real layer, the company positions its platform as an algorithm engineer’s “playground.” It continuously trains and deploys on G1, H1, Franka and wheeled bodies, then feeds real-world feedback back into the simulation pipeline.
17. Full-Stack Simulation Is Not a Module a Model Company Can Simply Build In-House
谢晨 believes simulation requires critical mass. Even Physical Intelligence, with world-class embodied-algorithm talent, spent a long time searching for a simulation lead. His explanation is that the company publicly downplayed simulation while internally depending heavily on it, making the positioning difficult for candidates to embrace.
Full-stack capability combines 4 areas: simulation technology, scaled engineering, real-to-sim and sim-to-real algorithms, and large-scale operations that research teams are usually reluctant to handle.
Physical information collection requires human participation, and quality verification cannot be fully automated. “People are the real standard for improving data quality.” This is also his explanation for why Scale’s data becomes more expensive as it sells more of it.
18. Meta’s Scale Deal Was a Purchase of AI-Era Raw-Material Capability
谢晨 says the Scale transaction implied a valuation of roughly $30B, while Meta actually paid about $15B. He sees it as more like an acqui-hire, focused on core talent rather than the complete team itself.
His analogy is that if Microsoft’s CEO had known 10 years ago how AI would develop, the most valuable acquisition would have been NVIDIA because compute is the key capability for large models. Looking 10 years forward, Meta’s missing basic input is data.
His core judgment is that “the ticket to entry for a future $10T company is mastering AI data.” Alexandr Wang brings not just data production, but also judgment, talent organization and customer relationships.
19. The Alexandr Wang Succession Theory Assumes He Gave Up 10x Upside
谢晨 goes further and speculates that if Zuckerberg retires in 10 years, “it will most likely be Alexandr Wang” who succeeds him. 张小珺 asks, “Really?” and 谢晨 explicitly acknowledges that the view is based on his observations of Wang since 2018.
His argument is that Scale could potentially have gone public. Based on his personal forecast, a rise from $30B to $300B in 3–5 years is not impossible. Wang’s acceptance of a transaction near the current valuation therefore does not look like a move for retirement or cash-out.
If Wang took Scale from $30B to $300B, the incremental value would be roughly $270B. If, as Chief AI Officer, he took Meta from $1–2T toward $10T, the potential impact would be $8–9T.
谢晨 describes Wang as “extremely aggressive, extremely pushy.” The succession call is not an established fact, but a high-conviction speculation inferred from Wang’s ambition, personal values and the role Meta has given him.
20. Meta’s Data and Talent Problems Explain Why It Needs an Outside Organizer
谢晨 says Meta was not among Scale’s top 3 customers. Scale reserves its most important data-production personnel for the highest-paying customers, meaning Meta did not get the best data and could not train the best models.
He also speculates that Meta’s internal AI organization suffers from insufficient intensity and efficiency problems, requiring an external leader to reorganize its talent.
AI talent acquisition depends on relationships. Wang has served customers including OpenAI, DeepMind and Meta, knows who the key people are and can recruit through existing trust. 谢晨 sees his rapid hiring after joining as evidence that the value is already being realized.
21. The Sim-to-Real Gap Will Not Disappear; It Can Only Be Measured and Compressed
谢晨 acknowledges that from first principles, “there will always be a gap” between synthetic and real data. The key is not to claim the gap has been eliminated, but to know at what magnitude the data becomes useful.
A synthetic-data company that only delivers data and leaves training to the customer cannot know where the gap lies. It must carry out algorithm deployment itself and continuously put simulation-trained results onto real robots.
Improvement must also happen in layers: real-to-sim increases parameter diversity and accuracy, simulation improves the solver, sim-to-real improves generalization and utility, and real deployment feeds the next iteration.
22. The Embodied Industry Will Split into Four Layers Rather Than One Company Owning Everything
谢晨 describes embodied AI as simultaneously inventing “the next-generation vehicle platform” and the L4 and L5 algorithms on top of it. Both are extremely difficult, making a full-stack winner unlikely.
The first layer is hardware, including robots, dexterous hands and data-collection equipment. The second is foundation models. The third combines software, hardware and deployment in specific verticals. The fourth is end-to-end infrastructure centered on simulation.
The layers are not strictly isolated, but each has a different responsibility: hardware provides the body, models provide the “brain,” vertical companies combine software and hardware for deployment, and simulation companies handle real-to-sim, simulation and sim-to-real.
23. Unitree Is Turning Its Body into a De Facto Standard through Academic Adoption
谢晨 sees Unitree as the leading example at the hardware layer. It first supplied robots to academic and research institutions worldwide, allowing top laboratories to publish papers on its platforms.
When graduates from those laboratories enter industry, they take their existing tools, assumptions and usage habits with them. Academic penetration then gradually hardens into an industry standard.
Complete robots are not the only products in this layer. Dexterous hands and data-collection hardware also matter. The core competition is not simply selling equipment, but making that equipment the default platform for algorithm research.
24. The Foundation-Model Layer Will Ultimately Require 100,000 GPUs, Data and Top Talent
谢晨 lists Physical Intelligence and Skild AI as independent model companies. The former publicly appears to rely mainly on real data, although 谢晨 emphasizes that it also wants to use simulation internally. The latter places greater emphasis on simulation, so the two have visibly different public methodologies.
Nvidia’s internal Isaac team and Google DeepMind’s embodied-AI group also belong to this layer, effectively serving as the robot’s “brain.”
He expects foundation models to “definitely use more than 100,000 GPUs,” along with massive datasets and the world’s best talent. The combination gives large technology companies a significant advantage at the model layer.
25. Physical Intelligence’s Contradiction Exposes the Scaling Bottleneck of the Real-Data Route
谢晨 sees PI as more like a top laboratory built by 2 professors, focused on finding the formula for embodied algorithms and data rather than immediately completing hardware, scaled deployment and commercialization.
Because simulation is more directly tied to scaling, it was not a high priority at the founding level. Internal researchers nevertheless understood its importance and continued demanding the capability, creating the tension of “not believing in it externally while hiring for it internally.”
谢晨 says PI rented apartments in San Francisco through Airbnb to collect data in real home environments. Even then, the number of apartments and the distribution of scenes remained far from sufficient. Expanding further would bring high rental and operating costs.
26. Good Data Is the Product of Scenario Diversity and High-Quality Demonstrations
谢晨’s formula is: “Good data is the product of two things.” The first is diverse scenarios spanning mathematics, programming, law, medicine and other domains. The second is demonstrations by high-quality humans.
He compares the imitation-learning stage to an ordinary teacher demonstrating an action. The more advanced RL stage is like the best teacher “setting questions” for the model to repeatedly attempt and solve.
Real-world teleoperation often remains stuck at the tabletop level: apples, bottles and kettles arranged on a table, or an IKEA-style mock apartment. The data volume can be large, but the scenes are narrow and generalization is limited.
Synthetic data also requires paying the cost of building scenes, but its “rent” is mainly compute. The same set of scenes can generate massive data in parallel, which is the core advantage for scaling the distribution.
27. Figure and Tesla Represent the High-Valuation Route of Vertical Software-Hardware Integration
谢晨 defines Figure as a vertical-deployment company: it will first enter automaker plants and may eventually expand into Amazon-style fulfillment centers. He believes the home-robot narrative contains “a lot of promotional material.”
Figure may fine-tune a foundation model and then complete sim-to-real using proprietary hardware. The episode puts its previous valuation at roughly $30B.
谢晨 also categorizes Tesla’s Omniverse as a vertically integrated software-and-hardware company and speculates that it may use xAI’s foundation models rather than Tesla’s own.
He also mentions The Bot Company and Dyna Robotics. The former targets the home; the latter starts with single tasks such as having a robot arm fold napkins in restaurants.
28. US-China Payment Structures Will Determine Company Boundaries
张小珺 asks whether software-hardware integration should be moved to China. 谢晨 believes the US still has an opening: some companies can leverage China’s industrial supply chain, while Figure and Tesla may retain limited full-stack R&D because of national manufacturing needs.
US users are willing to pay more for AI and autonomous driving. He notes that GPT Premium can cost $200 per month and that a ride from San Francisco to Silicon Valley can exceed $100, making specialized division of labor easier to sustain in the US.
Chinese foundation models are often free, and software monetization is weaker. Companies are therefore more likely to charge for hardware, concentrating the market in robot hardware and “big and complete” software-hardware businesses built around it.
The domestic “brain” may still come from teams at Galbot, ByteDance, Xiaomi and Li Auto, or companies may adopt overseas foundation models, fine-tune them with vertical real and synthetic data, and deploy them on their own hardware.
29. The Data Pyramid Is Converging the Real-Data and Simulation Camps
One public route relies mainly on real teleoperation data and imitation learning before adding a small amount of RL. Another centers on simulation. But 谢晨 observes that more leading teams are internally following the Data Pyramid.
The pyramid’s base is the largest volume of internet data, its middle is synthetic and simulation data, and its top is the smallest amount of real-robot data. The concept was proposed by 朱玉可.
谢晨 says some customers who publicly claim that “synthetic data is toxic” are in fact using synthetic data. Co-training real and synthetic data generally produces better deployment results.
The unresolved questions are the ratio, algorithm architecture and RL weighting. Most teams still use imitation learning to produce demos, while a small number of leaders are experimenting with RL fine-tuning for VLA models.
30. Nvidia’s Third Computer Elevates Simulation to a Strategic Layer
谢晨 recalls Huang Renxun saying internally, “NVIDIA is a simulation company.” Game graphics are already simulation for humans; physical AI simply changes the target from spectacle and fun to accuracy and real-world deployment.
Huang’s “three-computer problem” consists of the data-center computer, the edge physical-AI computer in cars and robots, and the simulation computer.
谢晨 infers from their equal status that Nvidia sees simulation as a market potentially comparable to the first 2, reaching the “trillion-to-$10T range.” This is his market judgment, not an industry statistic provided by the episode.
A complete simulation stack includes a physics solver, rendering, sim-ready assets and scenes, as well as APIs, frameworks and tools. “Simulation is not the same thing as a simulator.”
31. Isaac Sim and MuJoCo May Converge through the Open-Source Newton Ecosystem
Isaac Sim is built on the PhysX physics engine and Omniverse rendering pipeline, with the combination of physics and vision as its advantage. MuJoCo has a strong solver and API, but relatively weaker rendering.
谢晨 believes PhysX suffers from aging infrastructure. It was originally built for games, its code is CUDA-based, and optimizing and maintaining it for embodied AI is difficult.
MuJoCo originally ran on CPUs. The team once rewrote it for GPUs with JAX in the form of MJX, but 谢晨 says many developers dislike JAX, limiting ecosystem maintainability.
The next step may be for both sides to rewrite the MuJoCo solver in Warp, creating MuJoCo Warp and bringing it into NVIDIA Newton. Newton will be fully open source, and Guanglun could contribute code from its proprietary Warp modules to both sides.
32. Game Data Pushes Models toward Cross-Universe Generalization
谢晨 says language models have more or less exhausted internet data. “The largest data volume for GPT-5 must be synthetic data.” He also sees GPT itself as a synthetic-data machine that continuously generates natural language.
Embodied VLA models instead need 3D, robot-perspective, multi-sensor data with action trajectories, making their structure far more complex than pure text.
Game data may not be high quality enough, but its worlds and tasks are highly diverse, making it suitable for pretraining. Institutions such as DeepMind can have agents generate first-person data across different games and use it to improve model generalization.
Unpopular games with little commercial value may also become training assets. 谢晨 describes the end state as “cross-universe, cross-embodiment”: models must generalize not only across bodies, but across worlds.
33. Embodied AI Is Still at GPT-1, but a Scaling Law within 1–2 Years Is Possible
谢晨 believes the industry is still at the “GPT-1 stage” and has not found a stable Scaling Law. Adding more data and compute still does not guarantee continued improvement in model performance.
He uses Tesla FSD as an example. Early data expansion hit a bottleneck; only after the end-to-end architecture was connected did more data translate into stable gains. That was the critical “Scaling Law moment.”
Autonomous driving took roughly 10 years from its founding to FSD, but he believes embodied AI need not wait 5 or 10 years. PI, Nvidia’s Isaac team and DeepMind had talent close to Andrej Karpathy’s level from the beginning.
张小珺 notes that multimodal Scaling Law remains unresolved. 谢晨 responds with the visual capability of end-to-end autonomous driving and offers a conditional optimistic view: “If we’re optimistic,” finding an embodied Scaling Law within 1–2 years is “entirely possible.”
34. The Real Bubble Risk Is Financing Becoming the Team’s Reward Model
谢晨 does not believe that being early automatically means being a bubble. Talent, capital and the lessons from GPT on Transformers, data and compute scaling could significantly shorten the exploration cycle.
His concern is teams becoming overly focused on financing amounts and valuation. If the rewards come from speeches, demos and attracting investors, those behaviors become the organization’s “reward model,” while technical progress moves into second place.
He believes more strongly that a company should “serve someone else” from day 1, making customer payment the feedback on value and allowing real delivery to drive further progress.
谢晨 says Guanglun’s investors were mostly proactive inbound contacts, including Chentao Capital, Qi Lu, 王玉 of Infinite Fund and 华东. But the company wants financing to follow progress rather than become the main narrative.
35. The First Paying Customer Forced the Team to Unlearn Its Experience
Guanglun was founded in January 2023 and did not secure its first paying customer until roughly September or October of that year. The hardest part of the first 6 months was connecting to external interfaces, getting feedback and persuading the customer to keep buying.
谢晨 co-founded the company with a high-school classmate. The Cruise CEO once called him an “execution machine.” His belief is that “you have to start running first, and run fast enough to bring everyone around you along.”
During the transition into embodied AI, he initially believed autonomous-driving experience was enough to define the problem. He later found that the underlying understanding and real customer needs were completely different, so he approached the field as a beginner, interviewing undergraduates, PhDs and professors, speaking with dozens of people and rereading large numbers of papers.
Guanglun ultimately wants to become the global standard for embodied simulation, not merely a supplier. Simulation must achieve sim-to-real and remain accessible. “Simulation will definitely become the core key to the entire deployment of embodied AI”—that is 谢晨’s full conviction.