Videos of humanoid robots usually focus on physical prowess: backflips, running, carrying boxes, and screwing bolts. However, XPeng's Humanoid Robot Anthropomorphic Lab recently released a 5-minute and 45-second video where the robot barely moved its feet. It did only one thing—stood among three people and demonstrated "who to look at, whether it recognizes you, and which language to use to respond."
Therefore, the video released by XPeng this time actually tests four capabilities of the robot: identity establishment, comprehension and memory, embodied interaction, and social coexistence. It should be clarified first that this is an R&D demonstration, which does not equate to features users can buy tomorrow.
Below, Vehicle analyzes the product technology behind the video released by XPeng's Humanoid Robot Anthropomorphic Lab based on the video.
Part 1: A Robot That Doesn't Move Its Feet First Learns to Turn Around
The test background is that three testers take turns speaking and change positions midway. IRON turns its body toward the person speaking each time, and the turning amplitude varies: sometimes it just tilts its head, sometimes it turns its waist along with it, and sometimes it turns its entire body.
This sounds simple, but it actually involves two overlapping problems.
The first problem is finding the person. Humans can hear which direction a sound comes from in a noisy environment, and machines can too. The official solution is a 9-microphone array for sound source localization. But sound can only tell you the direction, not who it is—if three people are standing close together and the sound comes from the same area, no matter how accurate the localization is, it cannot distinguish the individuals. So there is a second step: using vision to look at the mouth. XPeng states it can "identify lip movements through a multimodal fusion mechanism to lock onto the speaker." Simply put, it watches whose mouth is moving to match the sound with the person.
The second problem is how to turn. Adjusting the body when talking to someone is subconscious for humans; a small angle means just turning the head, while a large angle requires moving the torso, rather than spinning in place every time. To replicate this sense of proportion, IRON must coordinate the head, waist, and legs together. The engineers in the video described it as "acoustics, vision, and motion control, with multiple systems seamlessly integrated."
The four capabilities and timecodes are all from the official viewing guide; the frames are extracted from the capability video, and the copyright belongs to XPeng Motors.
Part 2: Switching from Chinese to German Mid-Sentence Without a "Mode Switch"
Next, a tester switches from Chinese to German. IRON continues the conversation directly in German, without selecting the language first or giving any "switched to German" prompt tone. Finally, it even finishes the car model recommendation in German.
The technical explanation for this is divided into three layers: the base model used inherently possesses multilingual capabilities, followed by multilingual post-training, and then balancing the training data for different languages. Translated into everyday language, it means the robot "grows up" in several languages simultaneously from the very beginning, rather than listening to Chinese, translating it into an internal language first, and then translating it into German to speak. Humans growing up in a bilingual environment are in the same state: when the other party switches languages, you follow suit, without needing a switching action in between.
Part 3: Remembering You by Aligning Three Clues
Finally, this is the part I think is the most thoughtfully designed. The three testers changed positions and even changed their clothes, yet IRON still matched them with their previous conversations. Cards were displayed on the screen for each person: Xiaobai is a novice driver who wants a good-looking and easy-to-park car; Xiaohei is a family with three children, planning to buy a G9L for camping.
The act of changing clothes is crucial. If recognition relies solely on appearance, clothing color and standing position would become cheap judgment criteria. The demonstration would certainly look good, but it would fall apart in a different scenario. The official approach is to build files using three streams of information simultaneously: facial recognition, voiceprint, and dialogue content, ensuring that the same person points to the same memory at different times and in different states, and this memory will continuously accumulate and update. IRON said midway, "I remember you very clearly." The guide directly sets the frame of reference as humans: humans tend to forget over time and get confused when they know too many people.
Part 4: Q&A, with Answers Falling Back to Its Own Car Models
Finally, it is the shopping guide scenario. IRON answers questions by combining a local knowledge base and online retrieval, and then overlays the previously stored preferences to continue recommending. When the family with three children asks about space, it answers that the trunk of the G9L can hold camping gear for the whole family; the end credits read, "It can understand your preferences and proactively recommend."
Supporting this segment is local computing power: 3 self-developed Turing chips with an effective computing power of 2250 TOPS. The model runs on the robot itself, without the need to transmit video and dialogue to the cloud. This idea is consistent with its approach in cars; edge-side inference not only reduces latency but also alleviates privacy concerns.
These Four Segments Can Be Achieved Because the Cockpit Has Already Practiced Them
Seeing this, those familiar with the XPeng in-car system should find it somewhat familiar.
VLM is the abbreviation for Vision-Language Model. In layman's terms, it is a model that can both see images and understand speech. The Tianji AIOS 6.0 pushed by XPeng in January this year focuses precisely on this. The cameras in the cockpit observe the passengers' clothing and state, and the system proactively adjusts the temperature, turns on defogging, and adjusts the volume; preferences are stored in a local "personalized memory bank," with the official effective computing power on the cockpit side being about 750 TOPS. The XOS 6.3.0 in August took a step further, using the whole-vehicle brain Master Agent to connect the VLA responsible for driving and the VLM responsible for understanding people, so that a vague sentence can be broken down into tasks and assigned to various modules.
Comparing the two side by side, the capabilities match one-to-one: the cockpit looks at clothing and state, while the robot looks at lip movements and orientation; the cockpit can handle multi-turn dialogues and vague commands, while the robot can mix Chinese and German; the cockpit has a local memory bank to store preferences, while the robot has dialogue profiles bound to facial recognition and voiceprints; the cockpit proactively recommends based on local knowledge plus online connectivity, while the robot recommends car models based on family constraints.
What changes is the actuator. In the car, after understanding the user, what can be moved are the air conditioning, navigation, volume, and parking; on IRON, what can be moved becomes the head, waist, legs, and gestures. The intermediate layer of capabilities—"understanding people, remembering people, and answering people well"—is the same problem.
Therefore, the material for this video is easy to explain: what XPeng sent out first is the only layer that has been verified by mass production. According to the division of labor at the official launch event in November 2025, the big and small brains of IRON consist of three parts: VLT, VLA, and VLM. VLM manages dialogue interaction, VLA manages action and walking, and VLT manages thinking and decision-making. These four segments are basically the work of VLM, with only the turning motion involving motion control, and VLT never appeared independently throughout the process.
The cockpit side is based on public reporting; the official has not disclosed how the models on both sides share weights and data pipelines.
Also Need to See Clearly What the Video Didn't Provide
The demonstration is controlled: three people, a black studio set, no crowd interference, and the official hasn't announced how many times it was tested or how many times it failed. These four segments come from a single demonstration on the same timeline; only by testing it again in a noisy shopping mall can the conclusion be considered solid.
There are also several things unanswered: how long the memory can be stored, how many people it can remember, and how to delete it; exactly how much weight the cockpit model and the robot model share; as tour guiding and shopping guide are the first batch of scenarios, which stores they will enter and when; and the most crucial thing: when tasks that require movement, such as guiding the way and handing over items, can be tested. With 82 degrees of freedom, a catwalk gait, and hands with 22 degrees of freedom, actually none of these hard bones of robotics were tested.
Turning robots from remote-controlled toys into machines that can work autonomously has always been difficult when it comes to hands and feet. But this time, XPeng's video bypassed that part and first proved another thing: it can stand in a crowd like a human, knowing who to look at and remembering who you are.
Final Thoughts
Of course, this may not be XPeng shying away from the hard tasks to take the easy route, but rather the most efficient step at this stage—this layer of capability has already been deployed in the cockpits of hundreds of thousands of cars. However, the true watershed for robots is yet to come. Wait until the day VLT and VLA deliver their results, and then see if it can hand things to you; that will be the true Physical AI humanoid robot.