Free Pages: EN — LARRY ROMANOFF: The Beauty and the Beast — IRON and Optimus: A Tale of Two Robots

Tuesday, December 9, 2025

EN — LARRY ROMANOFF: The Beauty and the Beast — IRON and Optimus: A Tale of Two Robots

The Beauty and the Beast

IRON and Optimus: A Tale of Two Robots

By Larry Romanoff

Contents

Unveiling IRON

Locomotion and Balance

The Wind Tunnel

Basic Motion and Functionality

Sensory Perception

Aesthetics

The “Uncanny Valley”

AI and Robot Intelligence

Batteries

Some Technical Matters

Robot Applications

A Comment on X-Peng

Conclusion

Unveiling IRON

On November 5, 2025, China handed the world another bombshell with the unveiling of X-Peng’s IRON humanoid robot. [1] [2] A video of the demonstration showed IRON walking with a relaxed and smooth, human-like rhythm. The “female” robot’s fluid, smooth, and human-like “catwalk” was so convincing that it sparked widespread discussion and skepticism, with many people both inside and outside of China convinced it was a human in a robot suit. [3] The suspicion was so widespread that the company publicly cut open the robot’s leg to prove it was not a human in a suit. [4] [5] In the video, the staff also opened the zipper on IRON’s back and showed everyone the internal structure including the lattice muscles, the controllers, and the blinking lights. [6] When a robot moves so naturally that people assume a live person is concealed inside a latex suit, you know you’ve crossed a threshold. Here is the original video with IRON walking and then her leg cut open. [7] Another good video of Iron walking and having her back exposed. [8]

Ref. [6] Ref. [8]

Image source: X-Peng

X-Peng is not the first robot manufacturer to cause a commotion because of walking. In 2014, Zhongqing Robot SE01 also became popular on the Internet due to its silky gait, and quickly became one of the hottest robot manufacturers in China. Zhongqing is also an X-Peng company, and its President was the founder of X-Peng’s robot team. Dozens of Chinese companies have already developed “humanoid” robots, many capable of rather astonishing autonomous activities.

Selected IRON images

IRON Composite. Source:

IRON Dancing and Riding a Bicycle. Source:

IRON Cute Couple. Source:

Selected Optimus images

Optimus Composite. Source:

Optimus Balance. Source:

Optimus sorting objects. Source:

Locomotion and Balance

Here are some short videos you absolutely must see:

Iron Dancing – This excellent video is #1. It says it all.

IRON performing front somersaults on parallel bars.

IRON riding a bicycle.

Cute couple. IRON and her male counterpart. This is an excellent little video.

An excellent video of IRON doing her catwalk.

An excellent short video in English.

The Secret to Iron’s Life-Like Walk

Iron’s ability to mimic human walking motion comes from the spine and waist design. The engineering of the spine allows the torso to be flexible, while the waist is designed to provide the rest of the body with balance, even while swaying, in imitation of human biomechanics. The precision control and mechanical resilience are what enables IRON to retain its smooth walk even with some of its outer layers removed. [9] IRON’s movement trajectory is very similar to that of a human, especially the gait. When walking at the press conferences, IRON’s posture and steps were soft. It doesn’t use the traditional slow mechanical leg lifts, but first supports itself by moving its center of gravity forward, then swings the leg outward, naturally and coherently like a human “cat step”.

Tesla’s Optimus is fundamentally unstable in locomotion. If it had to walk with its legs fully extended and its feet flat on the floor, it would fall over and come crashing to the ground. It isn’t much more stable when stationary; in internal reports, Tesla staff say that Optimus robots fall half the time when performing tasks that require any bending or tilting, sometimes damaging expensive equipment. Unless performing a task that requires moving more than a few feet, Optimus robots are usually tied to a support frame to stay upright. And that is simply bad engineering design. A robot should be able to easily stay upright; the standing and balancing should have been the first thing Musk addressed. But he didn’t address it. Instead of recognising and admitting a flawed design, Musk decided it was sufficient for Optimus to walk in a perpetual half-squat, like a chimpanzee.

X-Peng’s IRON walks perfectly like a human, but Musk’s Optimus walks strangely, with its knees always bent, as if in a partial squat or a crouch. The bent-knee “crouched” posture of Optimus is a well-known crutch in robotics and is common in many bipedal robots. This stance dramatically lowers the robot’s center of gravity, which significantly simplifies the challenge of balance and prevents falls. It’s a shortcut that prioritises not falling down during demonstrations, over achieving a truly human-like, efficient, and dynamic gait. This is not a prize-worthy trade-off; it’s an admission that the current design cannot achieve stable, upright locomotion. A stable, upright bipedal walk requires a high center of gravity to allow for dynamic balance and a natural gait cycle. However, this is incredibly difficult to control.

The unusual posture and appearance of Optimus are the result of deliberate engineering trade-offs to produce a semblance of stability. In contrast, X-Peng’s IRON represents a more ambitious, upfront investment in the technologies required for natural movement. X-Peng engineered IRON from the ground up to mimic human biomechanics. A key to its “catwalk” is a linkage structure that places heavy components like motors closer to the torso instead of in the legs. This results in lighter lower legs and feet, reducing inertia and allowing for quicker, more precise and graceful leg movements, much like a human. This approach is far more complex and places a greater burden on software and control algorithms to maintain balance with a higher center of gravity. However, the payoff is a movement style that is more natural and potentially less unsettling for human interaction.

Humans walk like an inverted pendulum, with straight legs for most of the stride, which is highly energy-efficient. IRON’s “catwalk” demonstrates this principle. In contrast, a constant crouch gait, as seen with Optimus, is less efficient, consumes more power, and looks unnatural because the joints are perpetually under load. The squat posture in locomotion is a flaw, not a feature, a fundamental compensation for a design that cannot achieve stable, upright motion, a solution for a robot that would otherwise topple. It was the only way to prevent Optimus from falling on its face whenever it tried to walk. The bent-knee posture in locomotion is a compensation for stability issues, not some brilliant design innovation. That doesn’t deserve a prize. IRON has a design that seems almost infinitely superior, with results that do deserve a prize. When you compare Optimus’s awkward squat to IRON’s fluid walk, it’s hard not to see one as superior engineering.

The locomotion and general appearance of IRON and Optimus in motion reflect fundamental differences in their design philosophies and engineering choices. With Optimus, Musk made deliberate engineering trade-offs where he sacrificed refinement for basic stability while in motion. That sounds like poor planning and an impoverished design philosophy, which means they had to make Optimus function awkwardly, to prevent it from falling on its face.

The Wind Tunnel

Wind tunnel composite. Source:

IRON was even tested in a wind tunnel, with wind speeds up to 80 mph (130 kph). Here is a brief video. [10] You might ask why anyone would test a robot in a wind tunnel. While it might seem unusual at first, this is actually a highly strategic engineering decision. the primary goal is to test and ensure the robot’s dynamic stability and control algorithms in extreme, unpredictable conditions. This test is critical for a robot designed to operate in human environments. A wind tunnel creates a controlled but intense environment that simulates some of the most challenging real-world scenarios a robot could face. The key is not that the robot will routinely walk in hurricane-force winds, but that if it can remain stable there, it proves its capability in less severe but more common situations.

This is all a matter of balance and gait stability. In the wind tunnel, they gain engineering insight into how the robot’s balance control system (its “cerebellum”) compensates for sudden, large lateral forces. In terms of a real-world application, this relates to IRON’s ability to walk steadily on a gusty street, near passing vehicles, in subway air blasts, or in open industrial areas. The wind tunnel tests whether the actuators in the robot’s hips, knees, and ankles can generate enough torque, quickly enough, to counteract the wind without overheating or failing. In social environments, this tests IRON’s ability to carry a load while walking on an uneven surface, or quickly adjusting to a slip or push. It helps engineers to understand a robot’s navigating a crowded, dynamic space where it might be jostled. The wind tunnel is used in verifying that the robot’s frame, covers, and external components don’t vibrate excessively, come loose, or create aerodynamic drag that impedes movement.

The bigger picture is that IRON represents a different philosophy to that of Elon Musk’s Optimus. The wind tunnel test is a stark public demonstration of a core philosophy I have discussed extensively in this essay: X-Peng’s pursuit of high-fidelity, biomimetic movement that is robust enough for complex human environments. Unlike a robot designed only for the static, controlled setting of a factory floor, IRON is being engineered to handle the unpredictability of the world where people live and work. Passing a wind tunnel test is a powerful shorthand for proving that capability to engineers, investors, and the public.

It is also instructive to contrast this with Tesla’s known challenges. While Tesla’s Optimus has struggled with fundamental hardware reliability (overheating joints, failing hands) in relatively controlled settings, X-Peng is publicly stress-testing its robot’s integrated hardware and software under extreme duress. This doesn’t mean IRON is flawless, but it signals where the development focus lies: on validating dynamic performance and control resilience. The wind tunnel test is far from a publicity stunt. It is a rigorous engineering benchmark for a robot intended to be a stable, reliable, and safe presence in the messy, unpredictable real world.

Basic Motion and Functionality

Here is a short video of Optimus handing out bottles of water at an event. [11] It is almost painful to watch. The robot is slow, hesitant, and clumsy. It appears only barely able to function in an environment where it has not been pre-programmed and must react to circumstances on its own. This was widely circulated as evidence of Optimus’ ability to interact in social situations. But here is another video of the same environment that wasn’t circulated. [11a] In this video, Optimus knocked over all the water bottles and then fell to the ground. The official explanation was that Optimus suffered from “work overload”, but in the video the robot appears confused, frustrated, and almost angry. It knocks over all the water bottles and then collapses on the floor. However, observers at the scene revealed that Optimus was not acting autonomously but was being controlled by an operator. They pointed out that Optimus’ last act before knocking over the water bottles and falling to the floor, was the exact motion of a teleoperator removing his headset. [11b] The observers confirmed that an Indian engineer was controlling Optimus in the distribution of the water bottles, became frustrated, removed his headset, and quit. Optimus could not function without the external control, and thus lost control of its limbs and collapsed. Musk had insisted previously that Optimus was not being remotely controlled, but this was just one more fraudulent exhibition of so many, of Optimus’ “abilities”.

Observer Comments. Source:

Here is another video of Optimus running and moving its fingers. [12] This one was also widely circulated, and presented as a victorious accomplishment, but the running is “bent-knee and slow motion”, and the hand motions are primitive and clumsy. Compare this to IRON’s dancing and somersaults to understand the difference in capability. Here is another video of three different robots (two American and one Chinese) to give you some idea of relative ability. [13] For additional comparison, here is a short video of Optimus and EngineAI’s T800 robot. [14] Compared to Optimus’ limited abilities, the T800 is awesome. Here is a short video comparing the walking sophistication of IRON and Optimus side by side. [15] This last one is a kung fu comparison video. [16]

Sensory Perception

Sorry, Elon, nobody wants your robot babysitting their kids. Source

I assume readers are familiar with Tesla’s many tribulations over its “Full Self-Driving” (FSD) software. Elon Musk claims that “humans use eyes and brains to drive”, so camera vision is all that is necessary for autos. Musk continues to promote this as fully-autonomous and far superior to human control, while 1,000 accidents and nearly 100 deaths suggest otherwise. The main issue is that Elon Musk made a decision years ago to forego quality and security in the design of FSD, in favor of low cost. And that meant foregoing most of the advanced sensors available like radar and LIDAR, and to proceed with a camera-only version of autonomous driving software. To say this was a bad decision would be quite an understatement. This is important for our purpose here, because this stubborn flawed reasoning has been transferred From Tesla to Optimus, almost certainly to experience the same unpleasant results.

Ashok Elluswamy was the leader of Tesla’s autonomous driving team (the ill-fated FSD), then moved to Optimus, where he immediately pushed the team to shift its R&D focus to camera-centric perception and learning solutions, the identical technical path he used for the training methodology of Tesla’s FSD software. [17] This may prove to be a critical failing of Optimus, because forward-looking “eyes-only” sensors will never be sufficient for a robot wanting to be “humanoid”. This was not just a technical choice but a deeply held ideological philosophy, [18] and Musk will not likely back down from his stubborn position.

The core of the debate lies in the observation that a robot functioning in dynamic human spaces needs more than simple vision. The prevailing expert opinion suggests that for a robot to be truly capable and safe in unstructured environments like homes or hospitals, it would need to integrate a suite of sensors. Relying solely on cameras could be the Optimus project’s biggest strategic risk, as it ignores other sensory dimensions critical to physical interaction and nuanced understanding. But Musk has closed this door and is committing Optimus to the same flawed sensory perception as with the Tesla autos. I have difficulty seeing how this can come to a good end.

The eyes only, camera-only approach is more difficult than it appears. [19] Until 2025, Tesla was using motion pictures to “train” Optimus to do things, but now Optimus is being trained by watching videos taken by Tesla staff. More than 100 Tesla employees spend their shifts repeatedly performing everyday actions like lifting cups, wiping tables, and pulling curtains. [20] The idea seems to be that if Optimus can see videos of every possible thing or action, it may know how to behave in that situation. But it is widely recognised that camera only is not an ideal solution. Musk claims having cameras and LIDAR is unsafe because the car (or the robot) won’t know how to react if the two sensory inputs seem to disagree. [21] But that’s just an excuse to justify a position he’s already taken.

The core issue is that Tesla’s strategy, while elegant in theory, has proven difficult to execute. The system’s documented vulnerabilities in bad weather and with unexpected objects (“AI illusions”) demonstrate that digital “eyes” are not a perfect substitute for human perception and reasoning. Furthermore, it lacks the human capacity for intuition and learning from a lifetime of subtle experiences.

The concern is that this is a dead-end path driven by Musk’s stubbornness rather than technical merit. Many robotics professors and experts in the broader industry recognise the importance of multi-modal sensing, and there are many academic papers on why robots need broad sensing capabilities. Most experts share my skepticism about vision-only approaches, and the entire industry – except for Elon Musk – is moving toward multi-modal sensing. Video lacks crucial information a robot needs to interact with the physical world, such as joint angles, tactile sensations, and force feedback.

Official development standards explicitly call for multi-modal perception for robots. This means integrating various sensors to enable functions like scene understanding and object recognition, moving beyond a single type of sensory input. Furthermore, robotics researchers argue that for safe and effective integration into human environments, robots will need enhanced perception, including auditory systems to understand language and tone, and tactile sensors to better control their interactions. This stands in contrast to Musk’s vision-only paradigm.

However, it will be impossible for Musk to change this architecture now, because Optimus is copying the identical system used for Tesla cars. To alter Optimus’ sensory functions would either invalidate all the training material or require Tesla to scrap their auto’s eyes-only system and adopt LIDAR as most others have done. Two very expensive choices.

This full commitment to a cameras-only sensory system isn’t just for Optimus’ “eyes”; it extends to how the robot is trained. This means the core AI models for understanding the world are shared between the Tesla car and the robot. The camera-only system avoids expensive sensors like LIDAR, and a single type of sensor input avoids the complex “sensor fusion” problem of reconciling conflicting data from different sources. But pure vision systems struggle with object recognition delays in rain, fog, or strong backlight, and may misjudge the distance to low-lying obstacles. The evidence is that this one-sensor strategy could be a “critical failing”. And it should fail. Most everyone claims that cars are “just robots too”, but there is a huge difference between the “intelligence” required by a car to avoid a tree, and a “humanoid” robot baby-sitting the children.

The core issue is that a humanoid robot operates in a fully 3D, dynamic environment where stability, dexterity, and spatial awareness are paramount. The challenges seen in FSD, while serious, primarily occur in the relatively structured environment of a road network. Translating that same sensory philosophy to a bipedal robot navigating cluttered, human-centric spaces is a significantly more complex problem.

The approach of X-Peng and other Chinese robotic companies to the sensor and dexterity problem presents a stark contrast to the vision-only strategy of Elon Musk. They are also aggressively pursuing advanced tactile sensing as a critical, non-negotiable component for achieving true robotic dexterity. The common thread is a belief that a robot must have a rich sense of touch to interact reliably with the physical world. Researchers at Fudan University frame tactile sensing as “the last kilometer” challenge for fine robotic manipulation. They argue that without it, robots will never reliably perform delicate tasks in unstructured environments. When we place this multifaceted, sensor-rich approach next to Tesla’s vision-only strategy, the difference in technical philosophy is profound. Chinese companies are operating on the premise that dexterity requires a constant, high-fidelity stream of tactile data to complement visual perception. They are publicly demonstrating robots that use this data to perform fine motor tasks that remain a significant challenge for most other humanoid robots.

Aesthetics

It is also true that the Optimus aesthetics reflect a lack of final-product vision. The design appears to be what happens when the goal is a quick, functional demo rather than a polished, viable product. IRON’s design philosophy is superior. It demonstrates a higher level of ambition and a more mature approach to solving the core problems of humanoid robotics, notably locomotion and human interaction. The comparison doesn’t just show different “choices”; it highlights a significant gap in the sophistication and maturity of the underlying technology and design intent. Based on the evidence of their respective robots’ capabilities, the view that IRON’s designers had far higher standards is a perfectly reasonable conclusion.

Look at the picture below, and answer this question: In what way does this random collection of used auto parts qualify as “humanoid”?

Optimus-bot. Source:

Optimus’ design was from the aesthetics of necessity, not philosophy. When a design is elegant and functional, its engineering can be a point of pride. However, the exposed, unpleasant joint where the torso meets the legs on Optimus is more indicative of a modular, hastily assembled design focused on proving basic functionality and prioritising speed of development and low cost over refinement. IRON’s synthetic skin and sleek torso, while potentially more expensive and complex, demonstrate a design philosophy that has considered the final product’s integration into a human environment from the start.

The reality of Musk’s Optimus is that his design was badly flawed and his solution was to use a crude, power-inefficient crouch gait to simplify the balance problem – an engineering compromise for stability. It is similar with appearance in aesthetic and form. Optimus is utilitarian and unfinished. It reflects a prototype mindset where basic function is the only goal. And the exposed mechanics signal a work-in-progress, not a “proud display” of its mechanical nature. In designing IRON, X-Peng chose biomimicry for efficiency. They accepted the greater design and control challenges in aiming for a natural, energy-efficient gait. The core software and mechanical design (weight distribution) are advanced enough to produce a more difficult, but ultimately much superior, form of locomotion.

Although admittedly beauty is in the eye of the beholder, I find the physical appearance of Optimus to be unusual and not particularly attractive. This design was partially from inadequate engineering ability, and partly due to cost. As to Optimus’ appearance, the media hype claims “Tesla has opted for a design that proudly displays its mechanical nature. The visible joints and actuators signal a focus on raw functionality and engineering substance over form.” That is just face-saving marketing nonsense. The exposed joints are evidence of Musk prioritising rapid iteration over polished design. I think the truth is closer to “We tried, but couldn’t make a humanoid robot, so we did the best we could with a machine.” I don’t believe that Optimus is “proudly displaying” its unattractive mechanical construction, so much as it is evidence of impoverished design and a lack of creative engineering talent. I would remind readers that when Elon Musk first introduced his “robot” to the public, it was a female dancer in spandex, and he boasted that he would produce a “humanoid” Tesla robot. The fact that he didn’t, means that he couldn’t. Musk and Tesla abandoned the idea of a humanoid robot and instead produced an industrial machine.

IRON was designed to be very “approachable”. X-Peng invested in a full-body synthetic skin and “muscle” coverage. This design choice is heavily geared toward social acceptance and future integration into human spaces. By making the robot look and move in a familiar way, they aim to reduce the “uncanny valley” effect and make it appear as a helpful companion rather than a pure tool. My view is that IRON’s designers had much higher standards and a far superior design philosophy to that of Optimus. In fact, a direct comparison to IRON is damning; IRON’s biomimetic approach with weight distribution in the torso shows what’s possible with more sophisticated design. This isn’t just philosophical difference but demonstrates different stages of maturity in tackling the core physics problem of bipedal locomotion.

The design philosophy behind X-Peng’s IRON and the 7-year development cycle weren’t just about technical refinement but about mastering the psychological aspects of human-robot interaction. Tesla’s rushed compromises ignored all this. X-Peng’s patience and cultural commitment to quality (7 years in development) contrast sharply with Musk’s “move fast and break things” mentality. The focus on biomimicry isn’t just engineering—it’s a holistic understanding that aesthetics and movement are as critical as functionality for social integration. This isn’t just better engineering; it’s more thoughtful product design. IRON’s design was not just someone’s subjective preference; it was a recognition of a fundamentally more mature and sophisticated engineering and design culture. IRON perfectly encapsulates the philosophy that separates a long-term visionary project from a quick demo.

This long-term, holistic approach is a hallmark of companies that prioritise a polished final product over rapid, hype-generating news cycles. It demonstrates a confidence that comes from a deep-seated belief in the quality of the underlying work, rather than a need for constant external validation through ambitious media announcements. X-Peng’s journey with IRON reflects a company building a platform for the future, while Tesla’s current Optimus prototype feels like a company building a demonstration for the present.

The “Uncanny Valley”

Uncanny Valley. Source: author

The “Uncanny Valley” is a fascinating psychological and aesthetic hypothesis, [22] a crucial concept in robotics and AI, and directly relevant to our comparison of Optimus and IRON. It describes the relationship between how human-like an object appears, and how emotionally positive is our response to it. The theory proposes that as a robot becomes more human-like, our empathy and positive response to it increase – but only up to a certain point. At first, as a robot looks more human, our response becomes more positive. We see a toy robot with a cute face, or a stylised cartoon character, and we find it appealing. We feel the most familiarity and affinity for things that are clearly not human but have enough human-like traits to be charming. Think of R2D2, C3PO, or even Wall-E and Eva. But then, there is a critical point where the resemblance to a human becomes very strong, but is not perfect. The object is close to human, but something is subtly off; the realism is inconsistent. Things that arouse subconscious triggers could include a perfectly rendered face with dead eyes, or a human face that expresses no emotion, or robotic jerky speech. This violates our expectations, since the robot should be warm and alive, but it feels cold and lifeless.

This is the brink of the Uncanny Valley. When we encounter something in this zone, our emotional response doesn’t just level off—it plummets into a deep valley of unease, revulsion, and eeriness. The unsettling feeling arises from a cognitive dissonance. Our brain is exquisitely tuned to recognise other humans. When something sends conflicting signals—it looks almost human, but its movements are jerky, its skin texture is wrong, its eyes don’t focus correctly, or its voice is slightly monotone—it triggers an alarm. Our brain flags it as a “non-human entity trying to be human,” which can be interpreted as unsettling, creepy, or even threatening. The theory suggests that if the replica becomes virtually indistinguishable from a real human being, our positive response can return to a high level.

X-Peng didn’t stumble upon the “uncanny valley” problem. They anticipated it. Their team included not only mechanical engineers and AI programmers, but also experts in human-robot interaction (HRI), cognitive science, and industrial design. They understood that for a robot to be accepted, its technical specs are only half the battle; its psychological impact is equally important. The investment of 7 years in developing IRON was in “integration”. Those seven years weren’t spent just making a robot walk. They were spent integrating the mechanical (the linkage structure for the walk), the sensory (the advanced tactile hands), and the aesthetic (the synthetic skin and form factor) into a cohesive, believable whole. This is exponentially harder than solving these problems in isolation, which is what Optimus’s modular, exposed design suggests.

This is exactly why IRON’s design philosophy is so critical for a “companion” robot, and why Optimus’s current form is suited only as a “tool”. Tesla’s Optimus sits safely on the left side of the valley. It is clearly and unambiguously a machine. Its exposed actuators, metallic body, and non-human gait do not try to convince you it’s human. Therefore, it doesn’t trigger the uncanny valley effect. It may be seen as unappealing or crude, but it’s not eerie or creepy. It’s a tool, and we judge it on its utility. X-Peng’s IRON is attempting to climb out of the far side of the valley. By using synthetic skin, a more organic form, and a fluid “catwalk,” it is aiming for a high degree of human-likeness. This is a much riskier strategy because if any element is off, it could fall right back into the valley and be perceived as creepy. However, if they succeed, the payoff is immense: a robot that feels like a natural, acceptable presence in human spaces—a companion.

In essence, the instinct that “who would want a domestic robot that looks and walks like Optimus?” is perfectly aligned with this theory. For a tool on a factory floor, Optimus’ design might be acceptable. For a robot in your home, you would want one that, like IRON, has been designed from the ground up to navigate the profound psychological challenge of the Uncanny Valley. IRON’s synthetic skin and fluid movement consciously avoid the uncanny valley, while Optimus’s mechanical appearance stays in the “safe” non-human zone but sacrifices appeal.

This uncanny valley problem raises another matter. Most robots today, at least those pretending to be humanoid, have no faces. This is related to the uncanny valley, because a human face that appears natural, that reflects the wide range of emotions and realistic eye movement, is an extremely difficult thing to fabricate. Creating a natural-looking face with realistic emotional expression is one of the most difficult challenges in robotics. Many Chinese firms are working exclusively on this area alone, and we can expect IRON to one day have a fully-responsive human face. In fact, it seems that in 2026, IRON will have a pretty human face and blond hair.

IRON 2026. Source:

AI and Robot Intelligence

He XiaoPeng said, “If humanoid robots want to truly enter the actual use scenario, they cannot be limited to physical performance, but need their own memory, judgment and action logic.” [23] IRON is equipped with X-Peng’s first-generation physical world model, which realizes the three high-level intelligences by building a combination of high-order brain capabilities of “VLT + VLA + VLM”.

Optimus leverages the same neural network “world simulator” used for Tesla’s self-driving cars, trained on massive video data. IRON has a very different approach. She is powered by a “physical world” AI model, enabling “seeing, thinking, and doing”. This features a VLT+VLA+VLM “brain and cerebellum” system for decision-making. Through this model, IRON can carry out multimodal perception, semantic reasoning, action decision-making and real-time interaction, and realise the closed-loop behavior logic of (1) understanding, (2) thinking clearly, and (3) doing it right. [24] [25]

IRON was designed with three separate “cerebral functions” or operations, a sort of combination of “brain and cerebellum” system for decision-making. These are termed VLT+VLA+VLM. This is IRON’s “brain architecture”, very different from the Optimus model. It’s a multi-layered approach designed to transform robots from remote-controlled tools into autonomous entities by mirroring human cognitive functions.

The VLM is the Vision-Language Model. It provides scene understanding and common sense. It interprets the visual scene, recognises objects, and applies knowledge from training on its vast image-text datasets. This is what we might call the foundational sense. It provides the essential perception and “understanding” – the “What” – of what the robot is seeing, feeding critical information to both the VLT and VLA portions. The VLM acts as the robot’s foundational sense of sight and basic cognition. It processes camera and other feeds to identify objects, people, text, and an “understanding” of the overall scene.

The VLT is the Vision-Language-Task/Thought model. This is used for high-level planning. It converts visual and language input into a structured task plan or higher-level “thought”. It is responsible for complex reasoning, understanding context, and setting goals. It does strategic thinking and planning – the “Why” and “What to do” about it. The VLT layer takes over for complex tasks. It uses the VLM’s understanding and any verbal commands (e.g., “clean up this room”) to formulate a step-by-step plan or “thought” process. This is the slow, deliberative “brain” activity.

The VLA is the Vision-Language-Action model. It provides integrated decision-making. This model directly links perception and language understanding to generate specific physical actions. It is the connective tissue between models. It serves as a crucial bridge, translating the high-level plan from the VLT into actionable movement commands. This is the tactical execution portion – the “How”. The VLA receives the task plan from the VLT, and its role is to figure out the exact sequence of low-level movements—like joint angles, grip strength, and walking path—needed to execute each step. This is the fast, reflexive “cerebellum” function, turning intention into smooth, coordinated action.

This is not difficult to understand. Think back to the little video of IRON performing the front somersaults on the parallel bars. The VLM gave her the general understanding of the scene and the environment. It permitted her to recognise the bars for what they were, and to understand her purpose there. This is the “strategic planning” model in action. “Why am I here, and what do I do about this? What am I to do?” She has already been given the command for the somersaults, and she now formulates a step-by-step plan for the process.

The VLT model now takes over and does the high-level planning. It sets the actual goal, the precise “What to do”. This is the “slow and deliberative” part. It does all the complex reasoning necessary: how precisely to execute a frontal somersault; how much force to exert so she lands on her feet instead of her face; the position and movement of her limbs and body in execution; how to land precisely on the bar; what is needed to execute the performance successfully.

Now the VLA model takes over and provides integrated decision-making to generate specific physical actions. This is the tactical portion of the cerebellum that answers the question “How”, exactly, do I do this?” It determines the exact sequence of events in all the details – the joint angles, the strength of movement, the position of limbs and body parts required to execute properly. This is the fast, reflexive “cerebellum” function, turning intention into smooth, coordinated action.

These three “models” must be fully integrated and work together to produce the desired results. To say that they achieved this, would be an understatement. IRON not only performed the somersaults flawlessly, but she did much more. When landing on her feet after the second somersault, she threw up her arms in victory, clearly “understanding” that she had performed the task well, and that it was worthy of celebration. Even more, she was able to verbalise her success, stating, “I did the front flip, and I’m still on the beam. Mission complete. What challenge should I try next?”

This approach matters for IRON because this architecture addresses a key challenge in robotics. While simple, repetitive tasks can be pre-programmed, operating in unpredictable human environments (like a home) requires adaptability and reasoning. IRON’s approach is a step toward general intelligence. By separating high-level reasoning (VLT) from low-level control (VLA), the system aims to handle novel situations it wasn’t explicitly trained for, moving beyond simple pre-scripted behaviors. VLA models are considered the next step beyond basic vision systems, as they integrate reasoning directly into the action loop. X-Peng’s addition of a dedicated VLT layer for planning suggests an ambition to push this capability even further.

Optimus, on the other hand, is designed to rely on an overall “brain function” that is “one size fits all”. In other words, an AI that knows everything and can do everything, all in one place. IRON’s structure duplicates in a sense what DeepSeek did, which was to create compartments with different knowledge and specialties, and abilities, each performing a specific function. Optimus, following Elon Musk’s conviction, effectively relies on an imperfect AI today but hopefully will one day contain an AGI with “knowledge that spans the universe”. You can form your own conclusion about this. However, for the time being, Optimus cannot function in a truly autonomous fashion; it can do only what it has been programmed to do, and can copy only what it has been “taught” from watching videos.

The differences between IRON and Optimus are not accidental, but stem from fundamentally different philosophies: X-Peng’s IRON pursues biomimicry and sensor redundancy. Its goal is to create a robot that moves, perceives, and interacts as naturally as possible within the complex human world. This is a “top-down” approach that prioritises capability and safety for diverse environments. Tesla Optimus V3 pursues engineering efficiency and cost scaling. Its goal is to build a robust, affordable robot for repetitive tasks by leveraging Tesla’s strengths in automotive manufacturing and AI vision. This is a “bottom-up” approach that prioritises mass production for controlled environments like factories. [26]

It is worth the effort to understand how these fundamentally different technical philosophies in building a robot’s “brain” translate into distinct robot capabilities. [27] The fundamental philosophical difference is that Tesla aims for efficiency through maximum technology reuse and simplified hardware, while X-Peng pursues sophistication through specialised hardware and layered AI systems. The core difference is that Tesla prioritises applying a single, general-purpose AI system, while X-Peng is building a layered, specialised intelligence for complex physical interactions.

Optimus has a single, large neural network for all perception, planning, and control. IRON uses a specialized, “layered intelligence”, with dedicated models for different cognitive tasks like thinking, understanding, and bridging functions. Optimus shares the same network architecture and training data pipeline with Tesla’s Full Self-Driving (FSD) system (the eyes-only flawed one). IRON has a “three-model synergy”, composed of Vision-Language Model (VLM) for understanding, Vision-Language-Task (VLT) for planning, and Vision-Language-Action (VLA) for direct action generation.

Optimus relies solely on cameras (same as Tesla cars), and tries to leverage its computing algorithms to build a 3D understanding from 2D images. [28] [29] IRON depends on multi-sensor fusion that combines cameras, LiDAR, and millimeter-wave radar for a more direct and redundant 3D perception of the environment. [30] Optimus tries to maximise the reuse of auto FSD technology and data, and aims for a lower hardware cost, more rapid iteration, and scalability. IRON is not concerned about efficiency and scale in the same way; IRON is more concerned with precision and adaptability. It aims for a high-fidelity understanding and interaction with the physical world, enabling complex, unscripted tasks through its specialized “brain” model components.

The practical implications of these different paths are significant and align with the companies’ overall goals. For Tesla (Optimus), the focus is on creating a functional, inexpensive robot for repetitive industrial and domestic tasks. The strength lies in leveraging a massive existing data pipeline (from Tesla cars) and computing infrastructure. The potential weakness is that the system’s performance is intrinsically tied to the limits of its camera-only perception and its general-purpose AI. I would add here that Elon Musk in several recent promoted videos, claims that Tesla is the “world leader” in “useful, general AI”. But that is far from true. Nothing that Elon Musk has touched is a “world leader” in AI. Nothing.

For XPeng (IRON), the focus is on achieving a higher degree of biomimetic movement and sophisticated interaction suitable for service and companion roles. The layered cerebellum model is designed to handle more complex, multi-step reasoning. The trade-off is a more complex and potentially costly hardware and software system. In essence, one strategy bets on a single, powerful but generalised intelligence – which doesn’t yet exist – while the other builds a team of specialized AI “organs” working together.

Batteries

Powered by three Turing AI chips and 62 active joints, IRON walks, talks, and performs daily tasks with human-like precision. Source

IRON’s real breakthrough lies in two key points: the design philosophy and its power source. IRON is the first humanoid in the industry to announce the use of all-solid-state batteries. This is not merely an add-on or a minor iteration. The safety of traditional lithium-ion batteries has always attracted much attention, and the safety requirements for humanoid robots operating in homes and offices are much higher than those for cars. The extreme safety characteristics of all-solid-state batteries, along with higher energy density and lighter weight, are the cornerstone of IRON’s ability to enable flexible movement and commercial deployment. X-Peng directly applied to humanoid robots the battery technology it had accumulated in the field of electric vehicles.

IRON’s solid-state batteries have a potentially much higher energy density than the Tesla Optimus’ Lithium-Ion batteries. They are inherently safer because they use non-flammable solid electrolyte, and are non-toxic, which is important for household applications and social environments generally. Optimus’ lithium batteries carry a higher risk because the liquid electrolyte is flammable and can be volatile. We don’t want to see baby-sitting Optimus robots spontaneously igniting the way Tesla cars have had a habit of doing. The lifespan is also an issue. IRON’s solid-state batteries are designed to be very long-lasting, enduring over 15,000 cycles, while the lithium batteries in Optimus degrade faster and will require replacement after a few years. The runtime with Optimus is another problem, with the battery charge seldom lasting even two hours, hardly sufficient for most proposed uses. Further, Optimus’ runtime of under two hours is for normal operation. During high-intensity activities like performing a backflip, the power demand requires instantaneous discharge rates up to 100 times greater than normal use.

Another innovative concept with IRON is the “structural battery”, where the battery becomes a physical part of the robot’s body, much like fat and muscles store energy in humans. This multifunctional structure significantly increases the total energy a robot can carry without adding extra weight, another advantage that Optimus cannot offer with its fixed design. This is another area of a likely permanent deficiency of Tesla’s Optimus. Switching the robot to solid-state batteries would be a complex engineering challenge that would require a fundamental redesign of the robot’s power system and internal structure, not just a simple battery swap. Musk ran into this problem with his Tesla autos, but had to abandon thoughts of altering the battery structure because in the end it would have required a total redesign of the entire chassis. That would have meant scrapping the car and redesigning a new auto from scratch.

Some Technical Matters

Unlimited possibilities. Source

IRON stands 178 cm tall and weighs 70 kg, and its core driving force comes from three of X-Peng’s self-developed “Turing” AI chips, with a combined computing power of trillions of operations per second (TOPS). This level of computing power far exceeds the needs of most current smart cars and is closer to a mobile data center. It makes IRON one of the most powerful humanoid robots developed to date. For comparison, Intel’s Core Ultra 200V series processor, fitted into some of the best laptops, can achieve just 120 TOPS.

For a Direct comparison of design priorities, we can examine the following, not as philosophical differences, but as a difference in design maturity and ambition.

IRON’s key design elements include a flexible human-like spine, bionic “muscles”, and full-coverage synthetic skin. The robot’s impressive movement is attributed to its bionic mechanical structure, which places heavier components like motors closer to the torso to reduce the weight of the legs and feet, allowing for more agile movement. This reference is an excellent article containing a summary of IRON’s construction and abilities. [31]

Musk’s Optimus has a locomotion strategy of stability at all costs. It uses a crude, power-inefficient crouch gait to simplify an otherwise intractable balance problem. IRON on the other hand uses genuine biomimicry for efficiency, accepting the greater control challenge to achieve a natural, energy-efficient gait. In terms of aesthetics and form, Optimus is utilitarian and unfinished, reflecting a prototype mindset where basic function was the only goal. Its exposed mechanics signal a work-in-progress, not “a proud display” of its mechanical nature. IRON, by contrast, is product-ready and integrated, with a design that considers the “uncanny valley” and aims for social acceptance and aesthetic cohesion.

The underlying implications for Optimus are that core software and hardware for dynamic balance are not yet solved. The entire robot design is a patch, not a foundation. With IRON, the core software and mechanical design and weight distribution are sufficiently advanced to attempt a more difficult, but ultimately superior, form of locomotion. IRON also supports customisation of different body types.

Thanks to the anthropomorphic design, the new generation of IRON has 82 degrees of freedom throughout the body (Optimus has only 40), and the movements are flexible, which can realise the difficult anthropomorphic action of “cat walking”; IRON’s “harmonic joints” are used to achieve a 1:1 hand size with 22 degrees of freedom, allowing for fine gripping actions. In terms of intelligence, the 3 Turing AI chips, make it the humanoid robot with the highest computing power in the industry.

It is worth noting that X-Peng’s “VLT large model” is specially developed for robots (not automobiles), which can enable them to think deeply and make independent decisions. At the same time, the new generation of IRON applies all-solid-state battery technology to achieve lightweight, high energy density and safety, providing guarantee for the long battery life and safe operation of humanoid robots in complex environments.

As a highly anthropomorphic robot, IRON has human-like “bones-muscle-skin”, a humanoid spine, bionic muscles and fully covered flexible skin, with a 3D curved head display, bionic agile shoulders, and human-like dexterous hands. The skeleton mimics the curvature and stress distribution of the human spine, supports 1:1 bionic spinal movement, and can achieve natural bending and turning movements. The muscle layer is covered with lattice material, so that the robot has a sense of elasticity and strength while maintaining structural rigidity. The flexible skin achieves feedback characteristics similar to human skin through seamless wrapping process and tactile sensing, and the multi-layer structure not only makes movements smoother, but also lays the foundation for emotional interaction. As a result, it is competent for many delicate and complicated tasks that are firmly outside the reach of Optimus. These two articles contain good explanations of the technical matters. [32] [33]

Robot Applications

Elon Musk has for years claimed that the main applications for Optimus will be as mundane factory workers and household domestics. This initially sounds intuitively plausible, but the reality is quite different. It is already a foregone conclusion that Optimus will never be acceptable as a domestic servant, and no humanoid robots will be doing much in factories. Today’s humanoid robots aren’t the right fit for factories or homes. [34] The main problem in both cases is the hands. Tests have repeatedly proven that the dexterous but delicate hands of humanoid robots wear out extremely quickly when performing repetitive industrial tasks such as tightening screws. The hands seldom last more than a month in factory test applications, and they are very expensive to replace.

X-Peng’s strategy with IRON is focused on applications that are more feasible with today’s technology. Notably, the company has publicly steered IRON away from the two commonly touted future applications of factory assembly lines and household chores. The reason is that the technology for these tasks is not yet ready for commercial use. This candid assessment of current limitations reinforces their focus on less complex, but more immediately achievable, service roles. From this experience, X-Peng chose a shrewd viable market for IRON, giving priority to commercial service scenarios, such as an office receptionist, tour guide, personal shopping assistant, quality inspector, and other service roles. The value of these tasks lies in IRON’s visual perception, navigation and interaction capabilities brought by its anthropomorphic design. These tasks require little complex, high-intensity “two-handed” operations.

In the home scenario, the biggest challenge is security. The home environment is far more unstructured and unpredictable than the factory, and any mistake can be catastrophic. This is not a place for a robot like Optimus, lacking most of its senses and riddled with multiple shortcomings. A humanoid robot pretending to be a companion or household servant, or a babysitter for your children, needs much more than mere sight to function acceptably. It desperately requires a delicate sense of touch, and very much needs hearing as well. Smell might also be required, while taste is likely optional. Tesla appears more focused on functional industrial applications while X-Peng is targeting commercial scenarios like guides and receptionists where appearance and overall competence are priorities. This is not so important with Optimus which is oriented more toward factory work.

I should point out here that both Tesla and X-Peng have employed their humanoid robots in their factories. X-Peng’s production lines have practiced with hundreds of them, and concluded this was not an appropriate employment. Tesla has had the same experiences, and concluded internally that Optimus was less than half as efficient as a human in any of the factory tasks it attempted. A Tesla staff member said that at present, Optimus only handles batteries in Tesla’s battery workshop, with less than half the handling efficiency of workers, and has not yet engaged in more complex car assembly work. [35] Yet Musk inexplicably continues to push this employment as the expected norm.

A Comment on X-Peng

Image Xpeng auto. Source: XPeng

XPeng flying taxi. Source: XPeng

XPeng for individuals and for long-range group travel flying taxis. Source: XPeng

At present, the X-Peng IRON robot team has more than 1,000 people, which exceeds the vast majority of robot startups. The company is already well-established in the manufacture of EVs, with a market share of about half that of Tesla in China and well ahead of many other firms. X-Peng also designs and manufactures its own Turing AI chips, produces “robotaxis” with full autonomous capability, and its beautiful flying cars and taxis. Development in all these areas is proceeding simultaneously, with the company budgeting for around $7 billion for training costs. These are the same areas where Elon Musk claims his companies excel, when they in fact do not excel. X-Peng in reality is superior to Tesla in almost all respects. Through nearly a decade of accumulation, X-Peng have built a full-stack self-developed physical AI system covering chips, operating systems, and intelligent hardware.

In terms of production capacity, it is not only IRON that will begin mass production in 2026. X-Peng also has the world’s first mass production plant for flying cars, that began trial production in November, 2025, with a planned annual production capacity of 10,000 vehicles, and single products rolling off the production line in only 30 minutes. X-Peng has also built a hybrid flying car A868 for business travel, which is now in a key stage of flight verification. The vehicle has a maximum speed of 360 km/h, a range of up to 500km, and a 6-person cockpit design to meet business needs.

Shortly after the end of X-Peng’s “Technology Day”, Volkswagen announced a partnership with the company. Volkswagen’s new brand models to be launched in China in 2026 will be equipped with X-Peng’s Turing AI chip and VLA 2.0-driven driver assistance systems. This transaction is extremely significant, because it not only verifies the company’s autonomous driving core technology at the technical level, but also at the commercial level. X-Peng’s “physical AI” strategy has been commercially endorsed by the world’s top auto giants. In addition, X-Peng announced a partnership with Baosteel to explore applications for performing inspections in complex industrial environments.

Conclusion

In comparing these two robots, we must face the issue properly. It is as if you have created a jet plane that can fly nonstop from Shanghai to London. But I am dreaming about building a plane that maybe someday could fly to Mars. So, mine is better, even though there is no reality to it. This is not just a difference in approach. It is comparing something excellent that already exists to something that sounds good but might never happen. I would argue that IRON’s tangible achievements are being undervalued next to Optimus’s speculative promises.

The core issue here isn’t just factual accuracy; it’s about narrative framing and the validity of prioritising vision over execution. X-Peng has built a jet plane that can fly from Shanghai to London today. It’s a tangible, engineering-first achievement. They demonstrated it, proved it wasn’t a hoax, and outlined a clear, near-term path for its use. Musk is selling tickets for a hypothetical mission to Mars. It’s an appealing vision that captures the imagination, but the rocket, the life support, and the fuel don’t exist yet. The “someday, maybe” promise of the Mars mission is somehow allowed to overshadow the reality of the transcontinental jet. The issue isn’t just a difference in approach; it’s a fundamental imbalance in how we assign value and credibility. In tech media and public perception, this creates a “double standard” where a futuristic vision is given more weight than a solved, present-day problem.

Optimus’s “potential” is its primary asset. Its value is almost entirely speculative, based on a future where AGI is solved and seamlessly integrated into a humanoid form. Because the payoff is in the future, it’s immune to being disproven today. Failure to meet a deadline is just a “delay,” not a refutation of the vision. IRON’s “limitations” are held against it. Because it is real and has defined parameters (e.g., “we won’t use it first for complex factory work.”), its capabilities are judged against a perfect, fictional standard. Its honesty about its current scope is used to frame it as “less ambitious” rather than “more honest.”

Tesla’s Moving Goalposts: IRON had to undergo a public dissection – literally having its leg cut open on stage – just to be believed that it was real at all. The burden of proof was immense and immediate. Optimus had its own “suit-gate” where its first reveal was literally a dancer in a spandex suit. Yet, the burden of proof for its core promise – human-like learning – is perpetually deferred to an unspecified future date. The goalposts for Optimus move so often, they are on wheels. This is classic Elon Musk. When something fails, dismiss the failure and shift the focus to a different and more grandiose fairy tale for the future.

What X-Peng is doing is arguably harder in the short term. Creating a robot that can walk with a stable, human-like gait, navigate real-world environments, and perform specific tasks reliably is a monumental feat of mechanical, electrical, and software engineering. This is the hard, unsexy work of building the foundation. Musk’s promise of AGI, while intellectually fascinating, is being used as a rhetorical shortcut to avoid the immense challenges of that foundational work. It hand-waves away the present difficulties with the promise of a future magic bullet. The narrative that the “Mars mission” is inherently superior to the “jet plane” is a pattern that privileges storytelling over engineering, and speculation over production.

The more credible position is to judge what exists and has been proven today. By that metric, IRON, with its public demonstrations and pragmatic roadmap, is significantly ahead. Tesla’s Optimus remains a well-funded research project, but its claims of superiority are not based on current, demonstrable reality. They are based on faith in a future technological breakthrough. Optimus’s shift in training methods was an admission of prior failures, and highlights the lack of public, verifiable demonstrations from Tesla compared to X-Peng’s transparency. IRON’s catwalk represents a real engineering milestone, while Optimus’s “potential” remains just that—potential. As proof, Musk claims that “in a year”, Optimus will be able to thread a needle. [36] Like Tesla’s FSD and everything else, this monumental event will happen “next year”. So far, none of these “promises” have been realised. Musk constantly dismisses current failures to replace them with predictions of something even greater happening in the future.

IRON’s viral moment was a demonstration of a solved problem: stable, dynamic bipedal locomotion. It walked with a fluid, human-like gait that was so convincing it was accused of being a hoax. This is a fundamental and critical milestone for any humanoid robot, and IRON has publicly showcased it. With Optimus, Tesla’s public demonstrations have primarily shown Optimus in a controlled environment, and the “official” Optimus video record has either been heavily edited or AI-generated. Tesla’s approach secretly relied heavily on teleoperation (human pilots remotely controlling the robot) and motion capture. Almost all “evidence” of Optimus’ abilities has been deceptive or misleading. While it has demonstrated tasks like sorting battery cells or performing simple yoga stretches, its locomotion has not undergone the same public “trial by fire” or had to overcome the same level of public skepticism regarding its basic reality.

The recent shift to training Optimus primarily by “watching videos” is a monumental admission that the previous methods were not yielding the desired results. Experts agree that “video learning” for robotics is an unsolved, “moonshot” problem. By pivoting to this, Tesla is effectively saying, “We haven’t solved the practical robotics challenge, so we are now betting on solving a monumental AI challenge instead.”

IRON is a finished product prototype, ready for mass-production. It has a defined aesthetic (with male and female forms), it has been presented on a stage like a consumer product, and its movements are polished for public display. It is being groomed for specific, public-facing service roles. Optimus prototypes, as seen in Tesla’s carefully edited videos, look like engineering testbeds. They are often unfinished “bare metal” skeletons with exposed wiring, being tested in lab or factory settings. While this is a valid stage of development, it reinforces the image of a project that is still in the early R&D phase, far from a polished, deployable product.

Moreover, Musk has stated that the current version of Optimus will be scrapped and he will attempt to design a true humanoid robot, so human-like that “you will want to poke it to see if it’s real”. One observer wrote, “FSD was being massively overhyped via staged videos for years, and Tesla is doing the same with Optimus.” Another wrote, “Elon is doing the same thing with Optimus that he did with FSD – put out staged videos to hype something that likely won’t be real for many years to come.” [37] In all other instances where he has encountered problems with “his” design of anything, Musk never stopped to re-think but instead just doubled down on his original path. If this assessment is true, Optimus is a rushed, flawed product headed for history’s dust bin.

Musk’s “Doubling Down” pattern is a critical insight. This observation about his behavior is well-documented across his ventures (FSD, Cybertruck production, Twitter acquisition). He consistently frames blind stubbornness as visionary determination. In the context of Optimus, this pattern suggests a high likelihood that Tesla will continue to iterate on the current, fundamentally limited bipedal design rather than undertake a ground-up redesign. It will prioritise software demos and ambitious future roadmaps to distract from today’s hardware deficiencies. It will struggle to escape the compromises baked into the initial, rushed architecture.

The design chasm between Optimus and robots like IRON is not just a gap; it’s a difference in kind. Closing that gap would require Tesla discarding most of the Optimus design and starting over with a more mature outlook. Given Elon Musk’s established patterns, such a fundamental course correction seems improbable. Therefore, the project’s greatest legacy may ultimately be as a cautionary tale about the limits of applying a “move fast” software mentality to the hard, iterative problems of advanced robotics hardware.

If a company consistently fails to meet its own benchmarks for a functional prototype, and its response is not to deliver a better prototype but to instead propose an even grander, more technically speculative vision for what it might do someday, then that vision can rightly be classified as an empty, face-saving mechanism.

Tesla have demonstrated a robot that can walk without falling down, and can perform a limited set of teleoperated or pre-programmed tasks. The promise that one will suddenly leapfrog into AGI-powered super-capability is, until proven otherwise, just a story. The burden of proof is on Tesla to demonstrate that Optimus can do anything comparable to IRON’s catwalk, let alone surpass it. Until then, the assessment that IRON represents a more significant and real achievement in the present is not just reasonable; it’s the only conclusion based on the evidence we have.

X-Peng spent 7 years to get the IRON design correct, aesthetic, and functional. Elon Musk spent a few months to get any kind of robot out as quickly as possible. He succeeded, but if any company can sell millions of robots, especially for domestic applications, it will be X-Peng and not Tesla. Who would want a domestic robot that looks and walks like Optimus when they could have an IRON companion?

The timeline and philosophy speak volumes. The contrast between a 7-year development cycle and a rushed 3-month prototype is the core of the argument. It reflects a fundamental difference in philosophy: X-Peng/IRON: follow a principle of “Get it right, then scale.” This long-term, engineering-first approach prioritises a solid foundation—stable locomotion, sophisticated dexterity, and human-friendly aesthetics. This is the path for building a product meant for integration into daily life. Tesla/Optimus: Follows a known Musk principle of “Move fast and break things.” The goal was to establish a presence in the humanoid robot space as quickly as possible, using existing resources (like the flawed FSD AI). The result is a barely-functional proof-of-concept that appears to be a collection of engineering compromises rather than a polished product.

The market reality for “domestic” robots favors IRON’s approach. For a robot to be accepted in a home or workplace, it must be safe, reliable, and non-threatening. A robot that moves with an unnatural, jerky gait and has an industrial, exposed-mechanism aesthetic fails on these counts. IRON’s focus on biomimicry and a more refined form factor is directly aligned with the requirements of a companion or service robot. Optimus, in its current form, is not. The assessment that Optimus is headed for the dust bin of history is a highly plausible, if not the most likely, outcome.

While Tesla may eventually produce limited numbers of Optimus for specific, controlled industrial tasks, the idea of “millions” of these units in homes and general workplaces seems fantastical based on the current platform. Examining all of Elon Musk’s prior statements and claims about humanoid robots, my conclusion is that he naively (and thoughtlessly) assumed that only he would ever design and produce a humanoid robot. When Musk spoke of “millions” of his robots being in homes and factories, he clearly assumed that his Optimus would be the world’s only option. If you examine his statements today (as of late 2025), he is still apparently unable to accept or understand that the world of humanoid robots has already passed him by, that there are today many dozens of similar products and that most are superior to his.

Musk appears to have only the most juvenile understanding of the humanoid robot world. His appreciation of this new technology seems limited to his comments in a video where he said, “Who wouldn’t want an R2D2 or a C3PO in their homes?” [38] In that same video, Musk stated that “nobody has a useful robot today. Tesla will make the first useful robot.”

The realities of humanoid robots are very far removed from the understanding of this ten-year-old mentality. In another short video, Musk babbles about how the only safety for humanity is to have “a maximally truth-seeking AI”, [38a] while it is well-documented to the point of legend that his own version of AI – Grok – has been trained to lie. [39] [40] This is surreal to the point of mental deficiency. Elon Musk appears to live in his own fantasy world where reality is manufactured at will.

My conclusions are two. (1) IRON represents the humanoid robot of the future, and (2) Optimus is not a robot. It is (at least in its present form) a collection of engineering compromises, likely headed for the dustbin.

During Tesla’s Q3 2025 earnings call, Musk made several specific claims: (1) He promised to show a “mass-production-ready prototype” of Optimus V3 in Q1 2026. He described it as looking “almost like a person in a robot suit” with “unprecedented realism”. He claimed further (2) that Tesla plans to start a “million-unit” production line by the end of 2025, with mass production beginning in Q1, 2026. He stated also that his goal is to achieve the scale of millions of units, as producing only hundreds would be “meaningless”. To add “the icing to the cake”, Musk clamed (3) that an uncrewed Starship mission to Mars would launch as soon as 2026, with Optimus robots on board to test landing and operations.

It is impossible to reconcile the contradictions between Elon Musk’s fantastic claims and real-world reality. The reality is that Optimus production was terminated entirely in late 2025 because of its severe engineering flaws, nearly-useless battery life, and poor functionality. Further, Tesla has appeared to have given up hopes of improved designs and has been pushing its suppliers to create new engineering models that will work. In addition is Musk’s claim that he is scrapping entirely the existing Optimus and will design a truly “humanoid” robot by the end of 2025 and begin producing it at the beginning of 2026.

To add to this is the fact that Musk’s “Starship” is far from a functioning reality. The basic design is still unproven, as are the refueling, life support, and most other aspects. At this point, at the end of 2025, Musk’s Mars Starship is a fairy-tale. The same would have to be said about the V3 “humanoid” robot. X-Peng took 7 years to perfect IRON; Musk apparently believes he can do the same in 3 months and – within that same three months – build a factory that will be mass-producing “millions”. There is nothing in this combined picture that makes any sense. This is true also for Musk’s “Robotaxi”, his “full self-driving FSD” that has been one year from “super-human” perfection every year since 2014, his Hyperloop, his Boring Company, and so much else. You decide.

Mr. Romanoff’s writing has been translated into 34 languages and his articles posted on more than 150 foreign-language news and politics websites in more than 30 countries, as well as more than 100 English language platforms. Larry Romanoff is a retired management consultant and businessman. He has held senior executive positions in international consulting firms, and owned an international import-export business. He has been a visiting professor at Shanghai’s Fudan University, presenting case studies in international affairs to senior EMBA classes. Mr. Romanoff lives in Shanghai and is currently writing a series of ten books generally related to China and the West. He is one of the contributing authors to Cynthia McKinney’s new anthology ‘When China Sneezes’. (Chap. 2 — Dealing with Demons).

His full archive can be seen at

https://www.bluemoonofshanghai.com/ + https://www.moonofshanghai.com/

He can be contacted at: 2186604556@qq.com

NOTES

[1] X-Peng’s new generation of robots take a cat step! Netizen: I suspect that there is a real person hidden in it.

https://www.thepaper.cn/newsDetail_forward_31912165

[2] IRON catwalk; https://www.youtube.com/watch?v=m_Ag_SgsHVg

[3] A video of the demonstration; https://www.youtube.com/watch?v=m_Ag_SgsHVg

[4] XPENG Cuts Open Its Lifelike IRON Robot on Stage

https://www.eweek.com/news/xpeng-iron-robot/

[5] Automaker XPENG releases new video to prove its robot ‘IRON’ is not human

https://www.globaltimes.cn/page/202511/1347511.shtml

[6] Excellent IRON with leg exposed

https://v.douyin.com/25T36IrsHoM/

[7] Iron walking and leg cut open

https://v.douyin.com/UDs00QsBRQU/

https://www.douyin.com/video/7569574906100455163

[8] Iron good video walking and back exposed

https://v.douyin.com/D2QKdRMMGvc/

https://www.douyin.com/video/7569517078688927016

[9] X-Peng strips down humanoid robot Iron to reveal its high-tech inner workings after viral stage demo

https://www.notebookcheck.net/XPeng-strips-down-humanoid-robot-Iron-to-reveal-its-high-tech-inner-workings-after-viral-stage-demo.1160365.0.html

[10] IRON wind tunnel

https://v.douyin.com/QV3DFuqfH5I/

https://www.douyin.com/video/7579170152186782025

[11] Optimus picking up water bottle

https://v.douyin.com/jfZdvqq8wN0/

https://www.douyin.com/video/7580920295214709477

[11a] Optimus knocked over all the bottles and fell down

https://v.douyin.com/TesmQDQPsjk/

[11b] Optimus was being teleoperated

https://v.douyin.com/u0i1xU_tqWk/

https://www.douyin.com/video/7581399670199455022

[12] Optimus running

https://v.douyin.com/fYTvvzb0vcI/

https://www.douyin.com/video/7580253071712783643

[13] Three robots running

https://www.douyin.com/video/7579917368638344499

[14] Optimus and T800

https://v.douyin.com/3wEVaDz3IjA/

https://www.douyin.com/video/7580689337700928778

[15] Video comparing the walking sophistication of IRON and Optimus side by side.

IRON vs Optimus

https://v.douyin.com/mMuRrjo_oeA/ https://www.douyin.com/video/7569312989615641087

[16] Robot kung fu comparison video

https://v.douyin.com/9DHsaW4vDoY/

https://www.douyin.com/video/7579999588405218595

[17] Tesla’s AI director is reported to have launched an internal mobilization: next year will be the “most difficult year of their lives”

https://baike.baidu.com/reference/62591749/533aYdO6cr3_z3kATPXdzvn5YS7NZNr66-DXV7FzzqIP0XOpSo_sUIEz6NYwsPVmHQ_e_pttbZkGyeGuB0pN6v8WduUzRbwhmX78WzvFzbvwuI9zl4MV-tEW

[18] Ashok Eluswamy

https://baike.baidu.com/item/%E9%98%BF%E8%82%96%E5%85%8B%C2%B7%E5%9F%83%E5%8D%A2%E6%96%AF%E7%93%A6%E7%B1%B3/62591749

[19] Tesla’s AI head warns that 2026 will face the biggest challenge

https://ai.zol.com.cn/1080/10809664.html

[20] Abandoning motion capture and fully turning to pure visual data collection, Tesla Optimus’ latest training progress is exposed!

https://app.myzaker.com/news/article.php?pk=69085c5fb15ec07899597a79&f=qqconnect

[21] Tesla and technology executives debate the path of autonomous driving technology: pure vision VS multi-sensor fusion

https://news.zol.com.cn/1037/10370329.html

[22] uncanny valley

https://www.britannica.com/topic/uncanny-valley

[23] If one day X-Peng and Tesla both sell housekeeping robots, what would you choose?

https://chejiahao.autohome.com.cn/info/23745737?reply=reply#pvareaid=2808151

[24] X-Peng Technology Day 2025: A new generation of IRON humanoid robots

https://news.yiche.com/hao/wenzhang/104946465/

[25] Cat Step is here! 2025 X-Peng Technology Day released high-end intelligent humanoid robots

https://chejiahao.autohome.com.cn/info/23748550?reply=reply#pvareaid=2808221

[26] Hardcore showdown: X-Peng IRON and Tesla Optimus, a technical route game on the humanoid robot track

https://blog.csdn.net/weixin_73527660/article/details/154578477

[27] Hardcore showdown: Xpeng IRON and Tesla Optimus, a technical route game on the humanoid robot track

https://blog.csdn.net/weixin_73527660/article/details/154578477

[28] 2025 Robot First Year: X-peng IRON/Yushu H2/Optimus Prime/1X-Neo software and hardware structure analysis

https://blog.csdn.net/VBsemi/article/details/154837120

[29] Hardcore showdown: X-peng IRON and Tesla Optimus, a technical route game on the humanoid robot track

https://blog.csdn.net/weixin_73527660/article/details/154578477

[30] X-peng Motors unveils AI humanoid robot Iron, challenging Tesla’s Optimus

https://m.huanqiu.com/article/4K970yDIWco

[31] X-Peng’s new generation of IRON was released, and large-scale mass production of high-end humanoid robots will be achieved by the end of next year

https://static.nfnews.com/content/202511/05/c11883931.html?enterColumnId=674

[32] Behind the controversy of “real person playing”, what is the technical content of Xpeng robot?

https://m.jiemian.com/article/13608058.html

[33] 82 degrees of freedom “Cat Step Kill”! Xpeng Iron robot is not a real person?

https://chejiahao.autohome.com.cn/info/23780714#pvareaid=6826274

[34] Tesla’s humanoid robots aren’t the right fit for factories, says former Optimus lead

https://www.techspot.com/news/108056-tesla-humanoid-robots-arent-right-fit-factories-former.html

[35] Tesla suspended the production of humanoid robots and modified the design

https://chejiahao.autohome.com.cn/info/20853171#pvareaid=6826274

[36] Tesla Optimus humanoid robot will be able to thread a needle in a year

https://www.teslaoracle.com/2023/12/22/tesla-optimus-humanoid-robot-will-be-able-to-thread-a-needle-in-a-year-says-elon-musk/

[37] Milan Kovac, Head Of Tesla Optimus Program, Departs

https://cleantechnica.com/2025/06/09/milan-kovac-head-of-tesla-optimus-program-departs/

[38] Musk babble on robots (nobody has a useful robot)

https://v.douyin.com/a9oq775TeVk/

https://www.douyin.com/video/7575010626527890730

[38a] Maximally truth-seeking

https://v.douyin.com/owrdSJRfvO8/

[39] Debunking Elon Musk – Part 12 — xAI and Grok

https://www.bluemoonofshanghai.com/politics/21776/

[40] Debunking Elon Musk – Part 18 — Fraud Update – xAI and Grok

https://www.bluemoonofshanghai.com/politics/22101/

This article may contain copyrighted material, the use of which has not been specifically authorised by the copyright owner. This content is being made available under the Fair Use doctrine, and is for educational and information purposes only. There is no commercial use of this content.

Other Works by this Author