184: Chen Zhe’s “Embodied AI Quarterly 26Q3”: After Astra’s Shock, Will General Models Eat Robotics?

Summary

• The quarter’s biggest shock was GPT-6 Astra, which showed that general models can enter physical AI: frontier labs are inclined to solve AGI first and treat physical AI as a subproblem. Astra does more than break tasks down; it can directly output end-effector poses, which are converted into motor signals through inverse dynamics—a step previously thought to belong to embodied models. Open-source replications showed a significantly higher success rate than Claude; Manqi also cited a separate Robot Dojo comparison in which Astra outperformed some leading embodied models. “If you haven’t seen the miracle, you simply won’t believe it.”
• Chen Zhe sees Astra’s impact in 3 layers: directly generating policies, supervising and correcting VLA, and—most “hair-raising” of all—acting as a robotics researcher. In Galaxy General’s case, Astra directly controlled 10 Robot Dojo tasks with a 26% success rate; supervising π0.5’s outputs raised that to 48%. The third-layer vision is to have Astra build simulation environments in Blender, generate objects and their properties, and debug motion-control policies: “In the physical world, AI can train AI.” Chen believes physical RSI is coming and is a direction to watch in Q4; he calls it “the latest fundraising concept.”
• The valuation story for embodied-model companies is under pressure: the assumption that “physical AI is bigger than digital AI,” which supports high valuations for Physical Intelligence and Generalist, may not hold. Astra suggests that “general models are a more formidable force, and physical AI may just be a subset of general models.” VLA may still have a role, but not necessarily be delivered by standalone VLA companies; general models may also distill into on-device models. At the same time, the market could expand and deployment accelerate, so this is not necessarily bad news for every startup.
• Chen expects every leading general-model company to make moves in robotics: there are rumors around Kimi and Zhipu, Liang Wenfeng has talked about physical AGI, and teams such as ByteDance Seed and Tencent RoboticsX may also pivot toward a generative-AI approach. The compute gap remains vast: leading embodied-AI companies may have GPUs in the thousands, versus hundreds of thousands or even millions at top general-model companies—a gap of 2 to 3 orders of magnitude. A friend at a Frontier Lab also cautioned that conclusions drawn from its data volumes and compute “do not apply to the vast majority of startups.”
• World Labs and AMD have signed an $8.2B acquisition agreement structured as an all-stock deal, but the transaction has not closed. Chen sees it as a reference point for exit discussions across the sector; there are also rumors that Anthropic may acquire Physical Intelligence. China has “basically no decent acquisition examples”; one leading embodied-AI founder said the company had raised RMB3B in cash but was in no rush to spend it, and wanted to wait for the bubble to burst before buying teams cheaply.
• Unitree’s IPO surged and then plunged—from a market cap of more than RMB440B at the open to more than RMB170B—but Chen says he “doesn’t feel much,” while Astra has made him more confident in a RMB100B valuation. Revenue in the first half was RMB1.152B, up 48% year over year. He drew a parallel with Nvidia: “No matter how successful OpenAI gets, Nvidia still benefits.” Meanwhile, regulators are tightening scrutiny of embodied-AI companies waiting to list. Chen warns that doing things contrary to the industry’s development just to generate revenue gets the priorities backward—like forcing OpenAI to monetize through a harness in the GPT-1 era.
• The embodied ICL capabilities shown by GEN-1.5 and S1 could cut the post-training needed for deployment to minutes of demonstration video. Two surprising findings: out-of-distribution tasks may be learned from a single demonstration, and long sequences of combined skills—such as packing and moving objects—can also be taught through video demonstrations. A growing industry view is that architecture may not be decisive: given data of the same quality and scale, VLA and WAM could both produce equally good policy models.
• At the end of July, the FCC announced restrictions on some advanced robots manufactured overseas, covering robots weighing more than 2 kg with mobility, perception, connectivity, and control capabilities. Chen says the share of core components that must be sourced domestically “should be” 65%; companies including Boston Dynamics, which had previously supported restrictions, have also appealed or raised objections, though the transcript does not clearly identify Figure as an appellant. The near-term impact is limited: already-certified Unitree models such as the G1, R1, and H2 can continue to be sold without major hardware or software changes. Over time, the policy could push companies to build factories in the US and bring suppliers overseas, but China’s supply-chain advantage, built over many years, will be “hard to change through short-term restrictive policies.”

Deep dive

1. Five big stories in Q3: Astra leads, alongside M&A, the IPO, ICL, and FCC restrictions

• Chen Zhe’s ranking starts with GPT-6 Astra: “People saw the capabilities and potential of a general intelligence entering embodied intelligence—or physical intelligence.” It was the quarter’s biggest change. Second was Li Fei-Fei’s World Labs signing an $8.2B acquisition agreement with AMD, bringing the world-model concept, popular in recent years, into the conversation around large-scale M&A.
• The other 3 stories were Unitree’s listing and its impact on how the industry is valued; the strong in-context learning capabilities of new models represented by Generalist’s GEN-1.5 and Skild AI’s S1; and the FCC’s late-July announcement of new authorization restrictions on some advanced robots manufactured overseas, which could affect hardware R&D and supply chains in China and the US.

2. A paradigm shift: solve AGI first; physical AI is a subproblem

• Peter summed up his US trip in one sentence: “The frontier labs, represented by OpenAI, see physical AI as something to tackle after solving AGI—as a subproblem.” The traditional view was that language models lacked spatial and physical understanding, requiring dedicated physical-AI models. Astra has shown that general models may also have spatial reasoning and execution capabilities.
• The timeline goes back to early 2025, when Gemini Robotics launched both a VLA version and a spatial-reasoning version, ER. At the time, the industry focused more on VLA; in retrospect, Google may already have seen the ability of general models to understand complex spatial environments. In July this year, Anthropic used Opus to call robot interfaces and achieve capabilities previously thought to require embodied models. “Language models entered the physical world through tool use at first”—a new form of computer use.

3. A key step: Astra directly outputs end-effector poses, quickly replicated by open source

• The old setup had an LLM at the top breaking down tasks and an embodied model below executing them. Astra directly generates end-effector poses, which an inverse-dynamics module converts into motor-control signals. This was “previously thought to be something embodied models should do,” bringing some spatial understanding and action output down to a lower level.
• The path to discovery: OpenAI’s official blog did not mention robotics applications. After the open-source community implemented the pipeline in July’s “Claude Plays Robotics,” it “may have adapted it within a day or two” of Astra’s release; replications found a significantly higher success rate than Claude. Manqi added that the Anthropic paper also tested GPT-5.24, GPT-5.21, Kimi K2.6, Qwen 3.6, and other models. She also cited comparisons in “An Unexpected Robot Policy” and Robot Dojo, where Astra had a higher success rate than some leading embodied models.

4. IROS dinner survey: about one-third of experts think “there’s nothing left in robotics worth studying”

• At IROS 2026 in Pittsburgh, this was a contentious topic. Traditional robotics researchers still work on control, touch, and force analysis, and they varied in how strongly they felt Astra’s impact. At a dinner, CMU assistant professor Shi Guanya surveyed more than 100 researchers, using concepts from The Three-Body Problem to group them into factions such as the Adventists, the Salvationists, and the Toolists.
• The 2 extreme views: one side thinks language models can solve everything and robotics has nothing left worth studying; the other believes general models will never solve many problems in physics. Chen says the largest, most pragmatic group sees LLMs as tools for studying specific robotics problems. The group that thinks there is nothing left to study in robotics “may have been about one-third” of those present.
• On the criticism that LLMs are too large and slow for real-time response, Chen’s historical view is that the right approach can always be found to optimize them: “It’s hard to fight a Moore’s Law that keeps iterating.” Those who rule out this path on the basis of compute or speed “may all be disappointed.”

5. Valuation stories under pressure: the assumption that physical AI ≥ digital AI may not hold

• The extreme camp’s argument is that the ultra-high valuations of embodied-AI companies such as Physical Intelligence and Generalist, despite having no revenue or commercial flywheel, rest on the core assumption that “physical AI is bigger than digital AI.” If true, these companies could be like early OpenAI in 2015 or 2016, supporting valuations of tens of billions of dollars at a very early stage.
• Astra’s arrival “is telling people that general models are a more formidable force, and physical AI may just be a subset of general models.” If that assumption holds, it could upend the valuation frameworks and value-chain positions of many physical-AI companies. But Chen adds that general models could also bring physical AI forward and expand the market: “It’s not necessarily bad for every startup.”

6. Astra’s 3 levels of use: policy, supervisor, researcher

• The most direct use is to run Astra as a policy on the open-source embodied-task benchmark Robot Dojo, generating poses and then solving for joint actions. The second is to use a language model to supervise and correct an action model. In a hybrid-control case from Galaxy General, Astra directly controlled 10 Robot Dojo tasks with a 26% success rate; having Astra supervise π0.5’s outputs raised it to 48%.
• The third level is “the most hair-raising”: imagining Astra as an actual robotics researcher. Through tool calls, it could build simulation environments in Blender, generate objects and properties, then debug motion-control policies—independently modeling, training, and optimizing policies. “In the physical world, AI can train AI”: physical RSI.
• Early attempts at real-to-sim-to-real exist, but no particularly significant work has emerged yet; until now, most of the relevant resources have gone into VLA and World Action Model approaches. Asked by Manqi whether this was a startup concept, Chen called it “the latest fundraising concept.” He believes physical RSI is definitely coming and is a direction to watch in Q4.

7. Strengths and weaknesses: strong spatial reasoning and generalization, weak physical precision

• Chen says Astra excels at tasks requiring reasoning and spatial understanding, and generalizes better. Previously, a VLA could “completely lose it” if a cup changed color or position; Astra can handle much larger variations. It is less suited to environments requiring physical precision or execution, such as buttoning a shirt very precisely. VLA models have specialized teleoperation and post-training data for such actions, while language models may have relatively little of this data.
• That might seem to support a layered approach, but Chen stresses the main reason for layering is not that large models will never be able to do the job. It is that general-purpose robots will probably need real-time, high-speed compute on the device. Like Robotaxi, they can tolerate a second or two of latency for low-frequency decisions, while much of the compute has to happen locally. The key point: “VLA models may still have a place,” but not necessarily as products from standalone VLA companies; future GPT models could also be distilled into on-device models.

8. Where the capabilities come from: Blender artists, cross-modal data, and the compute gap

• Chen breaks Astra’s strengths into 3 areas: spatial understanding, complex task reasoning, and calling abstract robot interfaces (programming). Spatial capabilities “should all come from data.” He has heard that OpenAI may have hired many Blender artists as early as May to create modeling trajectories and label data, which could also explain Astra’s progress in generating Blender content. Once a proprietary test environment is built, Blender also makes it easy to verify results and measure performance.
• A broader explanation is: “If OpenAI’s goal is to build AGI, then it needs to learn and absorb data from every modality.” On whether pretraining used embodied data, such as first-person-view data, Chen says he has no definitive confirmation, but everything he has heard points to “definitely yes”—or, from a Scaling Law perspective, “there’s no reason not to.”
• The compute gap is structural: the leading embodied models may use GPUs in the thousands, while top general-model companies have hundreds of thousands or even millions. “The gap may be 2 to 3 orders of magnitude.”

9. Embodied-data companies: a market for this stage, with general AI companies as the biggest potential customers

• On whether demand for data vendors is a headwind or a tailwind, Chen’s view is: “Data companies serve whoever has the budget.” The biggest potential customers for embodied-data companies today may be general-AI companies. First-person-view or human-operation data is in short supply; once human-collected data builds up to a certain scale, it may also reinforce capabilities across other modalities.
• The more sobering view: “Every data business is tied to a particular stage.” Chen points to Scale AI: as models advance, data-service companies shift from facial recognition to autonomous driving to language models. “You always have to provide data on problems AI can’t solve today.”

10. VLA is closer to an instinct model: tens of billions versus a few trillion

• For some VLA researchers, VLA is closer to a VA model: language serves as scene description and prompt, not the key reasoning step. That is also why previous interviews criticized it as “a more advanced form of behavior cloning.” WAM uses a video-generation model as its backbone and general video data, potentially giving it much stronger generalization than VLA. But Chen still sees it as an action model: “It’s a generator, not an interpreter.”
• Manqi says humans seem to have had the ability to act before they became highly intelligent. Chen responds with examples such as insects and other lower animals: a small number of neurons can drive instinctive actions, while complex reasoning and task decomposition emerge as neuron counts and brain size grow. “Mainstream VLA and world-action models have tens of billions of parameters; mainstream language models are already at a few trillion, even above 10 trillion. That’s the difference between lower and higher animals.”

11. What startups should do: stop debating model routes, build the interface, keep the flywheel

• Chen’s view is: “Intelligence is a general problem. General models may solve intelligence first, then solve physical problems.” Going forward, embodied-AI companies should work with general-model companies as long-term complements. The first step is to stop arguing about whether VLA or WAM is the better route: in the face of superintelligence, differences in model architecture may not matter that much.
• The practical advice is to work out which capabilities general models can already provide and which still need to be trained or acquired in-house; build the interface—general models act as dispatchers, calling interfaces and capabilities in the physical world, and startups should make themselves into tools that can be called repeatedly; and retain their own flywheels, including touch, force sensing, process constraints, hardware reliability, recovery from failed tasks, and continuously accumulated task data.
• The intelligent-car analogy has 3 layers: automakers, Tier 1 suppliers, and core components. The middle layer must either move down into chips or, like Huawei’s smart-driving business, influence automakers from above. “If it just stays in the middle, it gets squeezed from both ends.” In vertical markets, companies can look to Harvey and OpenEvidence, using barriers to industry access or data compliance to build large businesses.

12. A decade of autonomous-driving benchmarks “vanishes in front of Astra”

• An embodied-AI founder with an autonomous-driving background told Chen that the corner-case benchmarks accumulated over more than a decade in autonomous driving—wheelchair users crossing the road, temporary construction cones—“vanish in front of Astra today.” Problems that human drivers often get wrong are easy for Astra, in his view. Chen followed up: “If this can happen in autonomous driving, why wouldn’t the same thing happen in embodied intelligence?”
• Chen’s net assessment is that the room for embodied-AI startups may narrow, but general-model progress will accelerate the deployment of physical AI and bring forward the emergence of large companies in the field: “The hardware for general-purpose robots is basically ready today. The remaining unsolved problems are really all intelligence problems.”

13. World Labs signs an acquisition agreement with AMD: an $8.2B all-stock deal

• World Labs is a classic World Model company. When Li Fei-Fei founded it, its mission was to build an all-modal spatial-intelligence model. Its Atlas model focuses on spatial reconstruction and viewpoint regeneration and can be used for games or robot simulation environments. This year, World Labs also acquired Synyx, founded by one of Li Fei-Fei’s former students, which works on robot-simulation validation and real-world testing. Together, they can form one part of a real-to-sim-to-real pipeline, rather than competing with VLA or World Action Model. As general models grow stronger, companies focused on spatial reconstruction or simulation environments may become more valuable.
• Chen’s third-party speculation on the deal’s rationale: Li Fei-Fei and World Labs still want to pursue frontier research, but an independent research lab without a clear business model needs substantial compute and resources; AMD can provide them. AMD, meanwhile, lacks a strong reputation and resources in AI. If the acquisition gives it a top physical-AI research team, “this is a very good deal for AMD.”
• The $8.2B figure may not reflect the business’s direct contribution, but “it shows what a leading company with a trillion-dollar market cap is willing to pay strategically for this.” Manqi added that the current proposal is an all-stock deal; Chen joked, “Even better—they don’t have to spend any cash.” The parties have signed an agreement, but the deal has not closed.

14. A window for selling out: Physical Intelligence acquisition rumors and China’s exit problem

• There have long been rumors that Anthropic might acquire Physical Intelligence, with a valuation and consideration even circulating. Chen believes the World Labs deal sets an anchor, and speculates that it is reminding similar companies that “in the short to medium term, the best way out may be to find a large-company acquirer as soon as possible.” For large companies, there may also be value only in being the first acquirer. There is a window to sell.
• The contrast in China is that there are “basically no decent acquisition examples.” Chen says he does not know how these companies can exit either, because no one wants to pay such high prices; people may think they can build it themselves or hire the team away. He cites an interview with a leading embodied-AI founder: the company has raised RMB3B in cash but will not spend much of it, and wants to wait for the bubble to burst before buying teams and companies at very low prices.
• Chen says US capital markets support model companies staying independent over the long term, and even unsuccessful attempts can often attract decent acquisition offers. That gives companies such as Physical Intelligence and Generalist the willingness to invest heavily in the directions they believe in. China’s world-model companies, by contrast, may pursue deployment and a commercial flywheel earlier.

15. Those who have seen the miracle and those who haven’t: a generational divide in how they think

• Chen’s takeaway from trips to Europe and the US: “People who have seen what OpenAI is doing think completely differently from those who haven’t. There’s a generational divide.” The Frontier Lab he saw believes the company is solving the problem of general intelligence, and everything else is a subproblem; he says this is a broad consensus across the company. “If you haven’t seen the miracle, you simply won’t believe it” (没有见到神迹的人就是不会信的).
• One example is Tesla’s launch of its end-to-end V12 model in early 2024. Many Chinese companies could not reproduce the results at the time; by mid-2024, some in autonomous driving were saying Tesla was creating smoke and mirrors and misleading the market. Chen says: “Maybe once you’ve actually solved a problem, your perspective on it changes completely.”
• It is also important to remember the blind spot: a friend at a Frontier Lab cautioned that reasoning based on massive data volumes and compute “may not apply to the vast majority of startups.” Chinese autonomous-driving companies can spend RMB4B to RMB5B a year training models; but if embodied AI is a subproblem of general intelligence, the scale is completely different. Manqi says they can only count on Liang Wenfeng to take on the challenge. Chen responds: “I think Liang and his team will do it.”

16. Chinese general-model companies may move faster into the field

• In September, there were rumors that Kimi and Zhipu would get into robotics; Liang Wenfeng has also said the next step toward AGI is physical AGI. Chen sees a clear trend: “All companies trying to solve general intelligence or AGI have the ability and the willingness to solve some physical-AI problems.” Leading large-model companies will make moves over the coming period. He expects Chinese companies to join in, since the data and compute invested in physical-space understanding could also benefit other areas.
• The more idealistic version is that anyone who genuinely wants to achieve spatial understanding at Astra’s level will have to train on relevant data; robots may even emerge naturally once that technical capability is in place. Zhipu had bought robotic arms a long time ago, but Chen thinks robotics was not a particularly important problem for any of these companies before Astra. The lesson from Q3: “We really have to understand and respect Scaling Law.”

17. Unitree’s IPO: “I don’t feel much” about the surge and slump, but the RMB100B case is stronger

• Last quarter, Chen put the figure at RMB100B. Unitree listed at a market cap of more than RMB60B, surged to more than RMB440B at the open, then fell back to more than RMB170B. He says he “doesn’t feel much” about the spike: new listings can see short-term speculation, which does not reflect long-term value. He is more focused on how long-term institutional investors view the company after the STAR Market lock-up expires next August.
• On fundamentals, Unitree’s revenue in the first half of 2026 was up 48% year over year (Manqi also gave the figure as 48%), with the share of humanoid-robot shipments rising sharply. Selling expenses were RMB164M; Manqi speculated that Spring Festival Gala marketing may have cost several tens of millions of renminbi, and Chen said “possibly.” He sees the company’s operating quality as best in class among China’s embodied-AI companies. After Astra, his view of Unitree has actually strengthened: accelerating general intelligence further raises the value of robust robot bodies and hardware companies.
• The Nvidia analogy is: “No matter how successful OpenAI gets, Nvidia still benefits.” Unitree can connect leading open-source embodied-model research and tools to its own robots, letting users deploy complete tasks at lower cost. Its 6B-parameter VLA model, released in September, focuses mainly on desktop manipulation and whole-body mobile manipulation. Its value lies in making it easier and cheaper to complete end-to-end tasks with Unitree hardware. Chen says traditional valuation methods struggle to price industry leaders; the key is scarcity, leadership, and future growth potential.

18. Tighter listing rules and “money games”: demanding revenue in the GPT-1 era would get things backward

• After Unitree, regulators tightened scrutiny of embodied-AI companies waiting to list, focusing on sustainable revenue, technological breakthroughs, and the need to go public. Looking back at the robotics boom of 2020 and 2021, Chen says the 2 leading delivery-robot companies, Keenon and Pudu, may each have burned RMB1B in a year. Subsidy-driven competition is hard to avoid in sectors awash with capital.
• Chen distinguishes 2 dangerous behaviors: using financial engineering to fabricate revenue, and “doing a lot of things that go against the industry’s development just to meet revenue requirements.” The analogy: if OpenAI had been forced to monetize in the GPT-1 or GPT-2 era, it might have spent resources on all kinds of strange harnesses and fine-tuning. In 2018 and 2019, Chinese companies built commercial applications such as smart customer service and smart healthcare on BERT and Google T5, but those efforts were ultimately shown to be at odds with the path to general AI and even ChatGPT. Manqi sums it up: “They picked up the sesame seeds and lost the watermelon.”
• What counts as real revenue? Chen says that if general intelligence is the main track, much of the real revenue will come from serving general-intelligence companies, through dexterous hands, data services, and robot bodies. The key is whether the business is positively correlated with general intelligence and Scaling Law. Using a large funding round to build traditional computer-vision systems that could already solve a dedicated industrial problem will not capture those benefits. He worries outside incentives could lead to “bad money driving out good.”

19. ICL: demonstration videos could replace hours or days of post-training

• GEN-1.5 and S1 systematically demonstrated embodied in-context learning: without changing model parameters, a task can be conveyed through a first- or third-person demonstration video. Deployment used to require collecting trajectories through real-robot teleoperation and then post-training for hours, days, or even a week. With ICL, Chen thinks a one-minute video may be enough for a model to learn a specific task.
• Two findings were unexpected. First, out-of-distribution tasks may be picked up from a single demonstration. Chen’s example: someone who has never tied shoelaces or used Velcro shoes will probably be able to after watching it once. Second, ICL works well for long sequences of combined skills: multistep tasks such as packing and moving involve recombining atomic skills, which a text prompt alone may not describe precisely.
• The 2 companies take different approaches. Generalist’s demonstration trajectories are shorter—the paper says 3 to 10 seconds—and it accepts raw robot-view trajectories that include joint information. Skild AI puts more emphasis on third-person and long-horizon video; its paper includes a 10-minute video example, highlighting long-horizon skill composition. Chen thinks ICL is attracting attention in China partly because the market puts more emphasis on commercial deployment.

20. Dyna 2.1 and the data pipeline: architecture may not be the key

• At the end of September, Dyna released 2.1. Its highlight may not be a model breakthrough, but a full workflow: a semi-humanoid robot, a whole-body control policy, and an engineering loop covering planning, execution, and feedback for laundry-room tasks. Dyna 2 had said it used more than 1 million hours of first-person-view data, but industry contacts say the Scaling effect is not especially apparent given that data volume. It is still unclear whether the model has hit a performance ceiling.
• Chen thinks network architecture may not be the most important factor; data mix, cleaning, and processing matter more. “Whether it’s VLA or World Action Model, with data of the same quality and scale, either could be trained into an equally good policy model.”
• Wang Qian, a founder of an autonomous-driving company, believes the data-processing pipeline needed to understand physical space may require the same work at OpenAI as at embodied-model companies, so the former may have no additional advantage. Chen agrees with the description of the pipeline, but not necessarily the conclusion: top large-model companies may have more advanced data-processing experience, compute, and tools. Today, many embodied-model companies still use GPT and Llama 3 as the “brains” when processing data. Manqi also points out that embodied companies may have no advantage in data pipelines and may even be behind; frontier labs may be using stronger internal models than the public versions.

21. Google’s “hot take”: the more you know about robotics, the more old experience may hold you back

• Google has worked on robotics for a long time, and has launched Gemini Robotics ER and Robotics 2 at the end of July, including ER 2 and an on-device model that runs locally. So why was OpenAI the first to deliver a shock? Chen points to Google’s organizational inertia and internal allocation of resources, and also thinks its foundation models were playing catch-up at the time.
• His own “hot take” is that the more a team knows about robotics and the more experience it has, the more that can get in the way of trying to make general models reach the same capabilities. People who have worked with hardware understand the complexity and uncertainty of the physical world, which may also make them resistant to new approaches. “The people who knew NLP best were also the ones who initially opposed LLMs most and had the least faith that LLMs could work.”
• When Chen worked at Google, the Everyday Robots lab was next door: “Back then, they were picking up trash in the office.” He thinks more than a decade of research experience can sometimes become historical baggage. In his observation, researchers working on robotics at Frontier Labs tend to be younger overall—typically 5 to 10 years younger than their Gemini Robotics counterparts. The researchers mentioned in the interview were born in 1999, 1980, and 2000, and some team members had just completed their PhDs.

22. Optimus out of focus, autonomy and robot intelligence “both related and different,” dexterous hands as training infrastructure

• Chen thinks Optimus is not getting much attention at the moment and is not a priority. In Q3, Tesla’s main focus was on deploying Robotaxi, and the teams share talent. Can autonomous driving transfer to robotics? His answer is “yes and no”: Tesla’s early world model was video-to-action, and driving is closer to instinctive reaction, with language playing a smaller role. Many embodied-AI companies have indeed adopted that approach. But autonomous-driving models are constrained by compute and data and may have only a few billion parameters; general models work at a different data scale and do more complex reasoning and understanding.
• Chen says the first-tier companies have not changed, but more companies are starting to work on dexterous hands. He sees them as essential equipment for AI training: more hardware means more scenarios can be tested and more real-robot validation can be done. “Suppose I have 1,000 hands and you have 100. Even if GPU compute is unlimited, I may train a model 10x faster than you.”

23. FCC restrictions: about 65% origin share, and the scope has also drawn objections

• On scope, the FCC handles compliance certification for relevant goods and electronic products; this time it brought advanced robots manufactured overseas under its certification jurisdiction. According to the public notice, robots weighing more than 2 kg with mobility, perception, connectivity, and control capabilities may be covered—not just humanoids and quadrupeds, but also some lawn mowers and robot vacuums. Chen says the share of core components that must come from the country of origin “should be 65%”; US brands could also be restricted if more than 65% of their core components are manufactured in China or a third country.
• During the FCC’s consultation, Boston Dynamics, US robot-vacuum startup Matic, Figure, and others expressed support. After the restrictions were announced, some US companies that had supported the ban also raised objections, arguing the scope was too broad. Boston Dynamics, which does a substantial amount of manufacturing in South Korea, is trying to negotiate an exemption or exception. Counting production in friendly countries toward domestic capacity at a certain rate is still only a proposal, not a final decision.
• The short-term impact is limited: products certified before the end of July can continue to be sold without major hardware or software changes. Unitree’s G1, R1, H2, and some robot-dog models are already certified; Chen thinks it would be difficult for the US market to find alternatives in the short term if Chinese brands were excluded entirely. In the long term, companies may need to build factories in the US and use domestically produced core components, potentially bringing suppliers along. Manqi compares this to Chinese automakers building factories overseas. Chen sees it as a long-term strategic goal; China’s supply-chain advantage, built through years of iteration, will be “hard to change through short-term restrictive policies.”

24. Q4 outlook: physical RSI, organizational shifts, and where the profits will go

• Chen expects more attempts at physical RSI around general models, including policy supervision, self-iteration, simulation environments and content generation, and using agentic approaches to orchestrate robot capabilities. He also expects Chinese general-model companies to move further into embodied research and the robotics market in Q4.
• Manqi mentions ByteDance Seed, Tencent RoboticsX, and other teams. Chen thinks these groups will adjust and move toward “building robotics guided by generative AI.” He also expects more ICL-related work, because ICL is easy to demonstrate and fits the Chinese market’s demand for deployment.
• Chen’s historical analogy for those resisting the new paradigm: 4 or 5 years ago, senior robotics researchers were still rejecting deep reinforcement learning as a control algorithm. By 2022 and 2023, deep reinforcement learning had transformed locomotion control in humanoid robots and robot dogs, replacing some traditional hand-coded rules and MPC methods. He thinks people can instinctively doubt general models too, but reality can be used to test them.
• The open question is: “If general intelligence keeps getting stronger, where will scarcity and profits in the robotics industry ultimately accrue?” In closing, Manqi adds that Astra is still slow at policy generation and its absolute success rate is low, so it cannot be deployed in real-world settings. She leans toward systems being layered in real environments in the future, but says what general models can already solve is enough to make people rethink the prospects for standalone embodied-AI companies. She also notes that demand for dexterous hands has recently been rising, and offers 2 long-term questions to watch: who provides the brain—embodied-AI startups or companies like OpenAI—and will OpenAI move toward Optimus-style full-stack vertical integration and build its own robot hardware?