Interview: Xiaomi's 陈龙 on Car-Robot Intelligence and Physical AI

Summary

• Xiaomi’s MiMo-Embodied is an embodied-brain framework that unifies autonomous-driving and robotics tasks, with its technical report saying the two capabilities already show mutually reinforcing synergies. The logic is that robotics data is scarce while assisted-driving data is more abundant. Multi-stage training and CoT connect the reasoning processes of the two tasks: robotics data teaches assisted driving spatial perception and spatial reasoning, while driving data teaches robots the laws of object motion. After reinforcement-learning post-training, the model reaches SOTA performance on both tasks.
• 陈龙 says an in-house foundation model is essential for automakers building VLA systems, and warns that training VLA on open-source models can be dangerous. Abstract or hazardous content mixed into internet pretraining could create a butterfly effect in critical assisted-driving decisions. Only an in-house capability—such as 罗福莉’s MiMo team—can properly filter data and feed corner cases back into pretraining. Major companies with both autonomous-driving and robotics businesses include Tesla, 小鹏 and Xiaomi, along with several smaller players.
• Xiaomi is exploring a next-generation architecture that combines XLA with latent-space reasoning: it does not require explicit language reasoning, but the latent space can be decoded back into language and future frames to expose the model’s thought process and provide guidance and supervision. Generating a long chain of text before producing an action would introduce “too much latency and too little control,” while producing no language at all could reduce the system to VA. Latent-space reasoning uses “the underlying machine language” for multi-step inference, making it faster and more comprehensive. User-facing functions already introduced include voice control of the car and parking-space-level navigation in malls, whose implementation 陈龙 considers best in the industry.
• Embodied-intelligence scaling laws are much harder to establish than those for language and vision: training loss follows a scaling law, but real-world deployment does not fully track the training-time relationship. The reason is that “embodied intelligence is fundamentally a systems-engineering problem.” Algorithms cannot perceive hardware and software differences such as tire friction, while the ability to recover from an erroneous state does not show up in open-loop training. The host also cited 高阳’s description of “very dirty data”; 陈龙 says that requires large-scale deployment scenarios. The next wave of scaling may come fastest from video self-supervision, with next-frame prediction implicitly teaching the laws of physics.
• The end-state, in his view, is that World Model and VLA will eventually converge: language is not everything, but it provides critical abstract reasoning capabilities. World Model primarily learns representations of underlying physical laws. 陈龙 says predicting how individual leaf-level pixels change is not the key decision problem; what matters is predicting more fundamental future states. VLA can bring common sense and accumulated human knowledge from language into the decision framework. Further out, supervised learning has a ceiling, and the end-state will require continual learning and eventually self-evolving systems.
• Lei Jun once spent an hour speaking with 陈龙 in person, and was won over by the end-state of people-car-home and the unification of cars and robots. They first connected in mid-2024, and 陈龙 took more than half a year to decide before joining around March 2025. Xiaomi has given him broad latitude for frontier exploration, with roughly 60% of his time allocated to research. On organizational design, he believes production and pre-research “should not be separated, but should gradually produce deployable results along the way.”
• Autonomous driving is entering deep water: the remaining takeover scenarios are fewer but harder, with the final “fraction of a percent” of edge cases becoming decisive. 陈龙 says talent is moving from autonomous driving into embodied AI, while China has advantages in data, the industrial chain and rapid iteration. Physical AI will ultimately depend on “testing your ideas in the real environment.” Over the next 3 years, he hopes to see L4 Robotaxi services achieve genuinely driverless operation and robots enter everyday life.

Deep dive

1. London’s Academic DNA: UCL, DeepMind and “Tuning Parameters by Making Coffee”

• 陈龙 earned his master’s degree from UCL—“DeepMind was essentially born at UCL.” His read is that London’s startup culture leans toward theoretical research. UCL, Oxford and Cambridge concentrate large pools of cutting-edge AI talent, with a character distinct from Silicon Valley. That academic DNA has carried through into his research taste.
• His signature analogy is worth preserving: “Making coffee is actually similar to training a neural network, because there are so many parameters to tune.” Grind size, water temperature, flow rate and pour speed all matter. “The key is that you can only make 1 or 2 cups of coffee a day, and each time you may adjust only 1 or 2 parameters. So it takes a while to reach the optimal settings, and every detail can have a major impact on the final result.”

2. Making the Last Possible TR35 Cut: From Biomedical Engineering into Autonomous Driving

• 陈龙 turns 35 this year—the last possible year to make the cut—and was selected for MIT TR35. The citation recognized breakthrough work on explainability in autonomous driving and the world’s first application of VLA to assisted driving. His cohort included researchers in brain-computer interfaces, nuclear fusion and biopharma: “Every industry has a small group of people using the latest technology to push it forward.”
• His path was not a straight computer-science track. He taught himself to code in primary school, deliberately avoided computer science at university, and studied biomedical engineering and medical imaging instead. He then went deep into computer vision—“an interface between computers and the real world.” After earning his PhD, he joined an AR startup that was acquired by a leading US ride-hailing company in less than 6 months, taking him into autonomous driving.
• Around 2018–2019, he launched a project to train neural networks on “low-quality” human-driving data collected from dashcams, learning human driving trajectories and validating a scaling law that was current at the time. The timing also coincided with the invention of Transformer in 2017, and the team adopted the new architecture.

3. The Original Sin of the Three-Stage Stack: A Wave Gets Lost in Transmission

• Autonomous driving 1.0 chained together 3 small models for perception, prediction and planning, with each stage outputting only structured information such as positions. “If the person in the car ahead waves you through, that information is completely lost in the segmented process.” A human would understand the gesture as yielding the right of way; the model would not see it.
• 陈龙 agreed that 诗婕’s analogy was “very vivid”: the three-stage stack is like 3 people passing water in a relay—“when the first person scoops up a basin and passes it to the second, only half a basin is left.” The host argued that the system was still fundamentally rule-based, and 陈龙 agreed: no human can enumerate every scenario in the physical world, which is why the system feels “very unlike a human” when driven.

4. The Aggressive End-to-End Bet: Faith in the Bitter Lesson and Moments of Doubt

• The founder of his former company was Cambridge academic Alex Kendall. Starting in 2018, the company pursued a fully data-driven, one-stage system: no object detection, no hand-defined rules, “just one neural network simulating human driving behavior, directly outputting action from video input.” It was “very aggressive” at the time, while the startup’s resources were, frankly, limited.
• The intellectual anchor was Richard Sutton’s The Bitter Lesson: “When people explore, they always want to add more things. … They reach a local optimum, but eventually you find that adding those things has little impact on the objective. In the end, the winning designs are the simplest structures.” 陈龙 also cited mathematician 许晨阳 in agreement: “The simpler, purer and more beautiful.”
• He openly admits to wavering: “A lot of the time it simply didn’t work, and nobody knew how to improve it.” He considered adding rules or reverting to a modular stack. Alex Kendall’s strong conviction that neural networks were the right solution gave him confidence. His own belief “was built up gradually.” After comparing Waymo’s more modular, absolute-safety-oriented approach, he got into his company’s car and found that the end-to-end system drove “very smoothly, very much like a human.”

5. FSD Validated the Direction for Them: “If He Has Done It, What Is Left for Us?”

• 陈龙 considers Tesla’s FSD to be, arguably, the first genuinely mass-produced pure-neural autonomous-driving system. His former team was “pretty happy” to see “our idea executed by a world-class heavyweight,” but there was also genuine anxiety: “If he has done it, what is left for us?” The answer was a different route: “We were actually closer to an open-source approach.”
• The former company left him with a methodology that maps closely onto today’s foundation-model playbook: establish a small baseline and clear evaluation metrics, then run large numbers of experiments and iterate quickly. It also taught him to resist “local ornamentation”—short-term gains that look prettier but are useless over the long term.

6. Lei Jun’s One-Hour Interview: From First-Generation Xiaomi User to Foundation-Model Lead

• At university, 陈龙 watched the launch presentation for Xiaomi’s first phone and immediately ordered one for RMB1,999. Compared with Nokia feature phones costing RMB4,000–5,000, it represented “roughly 2 months of living expenses.” In London, he watched the SU7 launch and was “extremely excited,” seeing a disruption on the scale of the first Xiaomi phone: “When SU7 launched, the new-energy vehicle market was already a red ocean, but Xiaomi still built a product with a very distinctive character.”
• Xiaomi’s recruiting team first approached him in mid-2024. “A lot of companies had contacted me,” but Xiaomi was the only one he wanted to speak with: “Lei Jun personally spent an hour talking with me. I really hadn’t expected that.” What ultimately persuaded him was the end-state: the people-car-home strategy, AI empowering every device, and the connections among cars, robots and smart devices. “It felt like I could accomplish more at Xiaomi, and make some contribution to the country.” He took more than half a year to decide and joined around March 2025.
• 叶航军 gave him the assignment: “Design a relatively frontier unified framework that brings autonomous driving and robotics together.” “That is exactly what I wanted to do, and one of the reasons I originally came back to China.”

7. Why He Put the World’s First VLA on a Car: Lingo and a Flat Organization

• In early 2023, the team began using ChatGPT to address autonomous-driving corner cases. “We found that it really did work,” and gradually turned it into Lingo, a portmanteau of language and go, for deployment in a real vehicle. It was “basically the first VLA model tested on a real car anywhere in the world.” At a time when even end-to-end was considered aggressive, adding a language model was “even more aggressive.”
• Looking back at “why you,” he points to 2 factors. First was the academic lineage: early investors included Ilya Sutskever, Yann LeCun and Peter Abbeel, “a very frontier group internationally in neural networks and AI.” Second was engineering and organization: the model was extremely simple—“1 person could automate training a model, deploy it to a vehicle, and experience the model they trained in the car the next day”—within an unusually flat structure where new ideas could be validated quickly.
• He acknowledges that the context has changed. His former company was more academic and paid less attention to scale and production; moving into Xiaomi’s public-company environment “changed quite a lot.” The team must now consider “not only the maturity of the technology, but also safety and user experience.”

8. The Cognitive Leap from End-to-End to VLA: An Ambulance Can Run a Red Light

• End-to-end is fully data-driven but remains a black box: “We don’t know what it is thinking internally, and it could produce very strange behavior.” It also does not truly understand the world. It knows that a red light means stop, “but it doesn’t know why you can’t run a red light.” If an ambulance is blocking the rear, traffic rules may allow a driver to edge through a red light under certain conditions. Such data may never appear in training—a highly long-tail scenario.
• VLA brings language capability into end-to-end systems. An LLM contributes world knowledge from internet pretraining and allows the system to articulate its judgment explicitly. 陈龙 agrees that language models can provide this layer of logical reasoning and explainability.

9. The Dream Team and the Hands-On Era for Technical Leaders

• The market calls Xiaomi’s autonomous-driving group a “dream team”—叶航军, 陈光, 王乃岩 and 陈龙—but 陈龙 says that is “a bit much.” The 4 are technical leaders in different areas: 陈光 leads L2, where VLA may be deployed first; former TuSimple chief scientist 王乃岩 leads L3; and 叶航军 oversees the whole effort while “actually using some frontier tools.” 陈龙 stresses that new industries need technical leaders who have “really trained these large models by hand,” so they understand what problems need solving.
• There is plenty of debate over the technical route, both inside and outside the company. Balancing short-term production with long-term exploration requires general-manager-level perspective: “Production and pre-research should not be separated. They should gradually lay eggs along the way, producing some production results during the pre-research process.” He spends roughly 60% of his own time on frontier work—reading papers, speaking with researchers and recruiting.

10. Two-Way Dependence Among the 90s-Born: 罗福莉’s Foundation Model and Why “Training VLA on Open-Source Models Is Very Dangerous”

• MiMo’s pretraining was mainly handled by 罗福莉’s team and later expanded into VL. 陈龙’s team does continued pretraining, strengthening spatial perception and spatial reasoning. He sends upstream the difficulties and data from downstream tasks, and arranges for interns to work with the other team. The host summarized the relationship—“You need its foundation model more, right?”—and 陈龙 answered yes, while the host noted that the foundation-model team also needs downstream feedback. The 2 90s-born leaders “often have meals together,” and “there are relatively few barriers.”
• His sharpest industry judgment in the episode is that an in-house foundation model is “very necessary” for assisted driving: “A lot of automakers today are actually using open-source models to train VLA. I feel that is very dangerous.” “Abstract content and dangerous things” mixed into internet pretraining could create a butterfly effect, affecting critical decisions in the final assisted-driving system and producing dangerous behavior. In-house capabilities are needed to filter the data and feed corner cases back into pretraining.

11. MiMo-Embodied: Putting Cars and Robots in the Same Classroom

• The host’s analogy won 陈龙’s approval: autonomous driving and robotics used to “attend separate small classes,” while Xiaomi is trying to combine them. The first reason is that robotics data is far scarcer than assisted-driving data, so the combination can improve robotics performance. The second is how humans learn: children acquire spatial perception and reasoning through interaction with objects, and “once humans have strong capabilities, learning to drive becomes easier and driving becomes safer.”
• There was no precedent for the bet: “When we first had the idea, we couldn’t be completely sure that the 2 could coexist.” Outdoor driving and indoor manipulation differ sharply in viewpoint, input and output modalities, and the industry had not validated the combination. The solution is multi-stage learning: strengthen driving first, then embodied intelligence, use CoT data to connect the 2, and finally apply reinforcement-learning post-training “to bring it to SOTA on both tasks.”
• CoT is the key unifying mechanism: “The modalities are very different, but the thought process can actually be unified.” Picking up a phone can be decomposed into recognition, localization and step-by-step execution; overtaking can be broken into similarly small reasoning segments. “These small tasks can produce synergies.”

12. How the Mutual Reinforcement Works: Robots Teach Space, Cars Teach Motion

• Robotics data consists of intense interaction—grasping, moving and engaging with all kinds of objects. It teaches assisted driving spatial perception and spatial reasoning. First-person assisted-driving data teaches robots the laws of object motion: “They learn the behavior of rigid bodies and the behavior of flexible objects from how different objects move.”
• Asked whether this was close to next-state prediction, 陈龙 confirmed that the model implicitly learns to judge distance and speed through the objective of “how you ultimately need to drive.” “Those capabilities may be relatively absent in robotics data.”

13. The World Model Definition Debate: “How a Leaf Changes Doesn’t Affect the Decision”

• Definitions remain diffuse. A narrow definition might be the one 杨立昆 has long advocated: predicting future states given an action. 李飞飞’s 3D-reconstruction work, simulators or 3D reconstruction in assisted driving, next-frame generation and world action models are all called World Models. “There is no particularly clear definition.” 陈龙 even considers language models World Models in essence because they “can use language to think about and predict the future.”
• His own definition leans toward generating the next state, but he does not see pixel-level generation as the point: “Predicting exactly how the leaves in this frame will change in the next frame has no effect on your final decision. … The most fundamental thing is not to generate a better-looking next-frame video, but to have a good representation for predicting the more fundamental future state.”
• The host noted that VLA and World Model attract so much attention because embodied intelligence and autonomous driving remain research problems that have not actually been solved. A new paradigm can become overhyped, raising expectations too far and crowding out other possibilities. The cycle resembles end-to-end’s progression from useless, to magical, to visibly flawed.

14. The Last Mile for VLA: Why Cars Are Moving Faster Than Robots

• The host cited 前嘉宾王星星’s public criticism that many companies claiming to build VLA are actually building VLM, with no action output reaching deployment. 陈龙 attributes the gap to data structure. Assisted driving has mature products, user-driving feedback and established perception-labeling systems that can quickly generate language data. Robotics teams mostly still have to collect their own data, and trail assisted driving by a wide margin in both cost and diversity.
• Cars also have a smaller action space: essentially 2D output, versus 3D multi-joint output for robots. Road structure and traffic rules make the driving environment more controllable. Xiaomi’s strategy is to let the faster track pull the slower one forward. The end-state is “one embodied brain controlling cars, robots and smart devices at home,” more like “the personal assistant we imagine for the future”—a group-level objective.
• He sees an early sign in the digital world through the “Lobster,” OpenClaw. Its agent generality is strong: it can operate a computer and program itself. At bottom, it is an LLM spawning sub-agents to assign tasks. “Eventually, when it enters the physical world, it will definitely work this way too.”

15. Sensor Philosophy: Mass Production Cannot Rely on Pure Vision Alone, and Sound Is a Modality

• 陈龙 believes “assisted driving can fundamentally be solved through pure vision,” but pure-vision systems in mass production “will definitely carry some risks” at the current stage. That is one of the distinctions he learned at Xiaomi between pre-research and production. Pre-research may achieve certain functions with vision, language and action alone; production needs more direct modalities to support decisions. Radar obtains true distance directly, while vision has distance ambiguity. Accurate range is critical for safe operation at high speed.
• The host added that sound matters too: “Ever since I felt less safe driving with headphones on, I realized that sound is actually quite important”—whether there is a car nearby, or whether someone is honking. When modalities conflict, 陈龙 says early fusion in an end-to-end system can let the system decide internally based on confidence and context. “We may trust vision a little more.”

16. Language Is Not Everything, but It Provides Important Reasoning: Don’t “Drive Blindfolded”

• The host raised Demis Hassabis’s position that language is not everything. 陈龙 agrees that language loses information and cannot explain everything, but calls it “a relatively abstract and powerful modality.” A few words can describe the essence of a scene, while the absence of a fixed time constraint also helps with long-term memory.
• The host added that much of humanity’s accumulated knowledge is compressed into language, while video has a shorter history of accumulation and a smaller relative scale. 陈龙 distinguishes a common misconception about VLA: it is not V converted into L and then into A. “That would be like driving blindfolded while your passenger tells you what is happening and you make the decision.” It is instead L added to the VA modalities: L reasons on top of V, and V and L jointly produce the action.

17. XLA and Latent-Space Reasoning: It Does Not Think in Language, but It Can Be Decoded Back into Language

• X stands for any modality and any embodiment—not only assisted-driving data, but also robotics data and internet image-text data. XLA is “the next generation of cognition-driven models.” It does not use explicit language reasoning: generating a long passage before producing an action would mean “too much latency and too little control,” while claiming not to output language at all means “you have not used reasoning capability; it may simply be VA.”
• Latent-space reasoning “is equivalent to using the underlying machine language for multi-step reasoning.” Latency is controllable, and because no translation is required, the information content is greater: “It not only thinks faster, but more comprehensively.” The latent space can be decoded into language and future frames, making it possible to understand “how it is actually thinking” and to provide guidance and supervision. He read 田渊栋’s paper on latent-space reasoning as soon as it came out last year and “strongly agreed” with it: “Humans do not need to think through every word internally.” Whether a narrow road is passable can be evaluated by mentally simulating the scene.
• On the user side, the gains include safer reasoning through complex scenarios—smoother routing around large obstacles and fewer takeovers—voice control of the car, with 小爱同学 handling acceleration, deceleration and lane changes, and parking-space-level navigation in malls that guides the driver to the available space closest to the elevator for the destination store. The underlying architecture remains a one-stage XLA system built on Transformer. Implementations differ across companies, and 陈龙 considers Xiaomi’s “effectively the best in the industry.”

18. Why Embodied Scaling Laws Are Hard: Systems Engineering and Dirty Data

• He has observed the gap firsthand: adding data does reduce loss according to a scaling law, “but when you deploy it in the real world, performance does not fully follow the training-time scaling law.” The core issue is that “embodied intelligence is fundamentally a systems-engineering problem.” “The friction between the tires and the ground is different, and your algorithm has no way to perceive that.” Errors are inevitable, while recovery from an erroneous state cannot be learned from the perfect trajectories of open-loop training.
• The host cited 高阳’s view that embodied intelligence does not need high-quality data, but rather “some very dirty data.” 陈龙 agreed, with the prerequisite of large-scale deployment. He believes the fastest source of an embodied scaling law may be video self-supervision: predicting the next frame implicitly requires learning object position, depth, speed and attributes—“where a pedestrian will be next second, how fast they can run, whether they are a child or a young person”—without labels and by implicitly learning physical laws.

19. The Ceiling of Supervised Learning: From Reinforcement Learning to Self-Evolving Systems

• 陈龙 says supervised learning “promoted the development of deep learning,” but also constrained the learning paradigm. Once an offline dataset is used for training, the system cannot iterate continuously, and its ability to generalize to OOD scenarios is limited: “You are confined to the world of the past.” To learn something new, “you have to do pretraining all over again, which is very cumbersome.” The end-state requires continual learning—“repeatedly practicing, correcting and fixing your mistakes in the real world, like humans.”
• Reinforcement learning “has solved some problems, but fundamentally still requires a more advanced learning paradigm,” such as self-evolving systems. Asked whether AI could explore freely without a prescribed reward, 陈龙 said it could also explore in the real environment. He does not rule out machines eventually abandoning human paradigms, but along the current path, “a self-iterating paradigm may be more promising.”
• The current mix resembles the LLM stack: imitation learning accounts for more of the system, and the scaling law appears there; reinforcement learning elicits capabilities and aligns with human preferences, while “the reward signal may be relatively sparse.” All of this depends on Transformer as a scalable network.

20. Deep Water and the End-State of Physical AI: “There Is Always a Chinese Person Smarter and More Hardworking Than You”

• The deep-water phase of autonomous driving has 2 meanings. First, it increasingly resembles a technology product, competing on intelligence and advanced features. Second, takeovers are becoming fewer but harder: “Can you solve the final fraction of a percent of difficult scenarios?” 陈龙 favors integrating World Model into the VLA framework, so the system can use both “the representation of underlying physical laws from World Model” and “the wisdom crystallized in human language” for reasoning. Even if VLA temporarily loses the World Model hype cycle, “the framework of cognition, reasoning and decision-making is not wrong.”
• The host cited a chart from 黄仁勋’s GTC speech placing Physical AI at the highest and furthest point after generative AI and Agentic AI. 陈龙 says the challenges include limited data, the inability to trial and error in the real world, and an academic field that has spent most of its time on supervised learning, with relatively few researchers exploring online learning. Hardware is another constraint, while “the hardware industrial chain’s future advantage will definitely be in China.” He used an overseas saying to explain his return: “There’s always a Chinese person smarter than you and working harder than you.”
• Asked what ultimately separates the teams, he declined to make a strong prediction but offered one standard: “You have to experiment in the real environment and validate your ideas.” Autonomous driving has produced large volumes of papers on virtual benchmarks, but performance in actual use may be much weaker. Many methods, he says, “ultimately turned out not to work particularly well.”

21. Talent, Consensus and the Three-Year Promise: Teaching a Robot to Make Better Coffee

• Autonomous-driving talent is moving into embodied-AI startups, where there are many new companies and relatively high pay. But both fields are embodied intelligence, and the transition is fast. A unified car-plus-robotics paradigm helps with recruiting: anyone who has worked in 2 of the 3 areas—large models, autonomous driving and automotive intelligence—is a fit. 陈龙 considers Xiaomi’s “co-creation and consensus” both the hardest and most necessary part: “If you start doing something without consensus, and the people below do not believe in it, they may simply give up when they encounter difficulties.” The car-computer integration effort overcame substantial resistance after consensus formed.
• The host asked a pointed question: if he had been turned down 3 times by Lei Jun, would there be a fourth attempt? 陈龙 replied, “I would definitely keep pushing. … I may not be able to completely persuade Lei Jun, but I will definitely do my best to make the case.” His observation of Lei Jun is that he personally engages with frontier talent; while preparing launch presentations, he even uses his own computer to edit scripts and train him for speeches. He is “very tolerant” with scientists, while still requiring milestone deliverables and regular reviews.
• His 3-year expectations are for L4 Robotaxi services to achieve genuinely driverless operation and for robots to enter daily life—to “make coffee for us.” The host joked that this would take away the fun. 陈龙 replied: “But we can teach the robot how to make better coffee.” The scene he imagined as a child while watching Knight Rider “is actually already very close.”