Chinese electric-vehicle maker Xpeng says its humanoid robot has entered mass production, with company representatives claiming it will be the first such machine to walk itself off the assembly line under its own power. The robot, called IRON, is being built through Xpeng’s existing manufacturing facilities, though the company has not disclosed a production target or unit count.
Xpeng representatives detailed the milestone to reporters in Berlin, describing IRON as a “high-level, general-purpose” machine capable of receiving instructions, planning a strategy, navigating its surroundings and avoiding obstacles without repeated trial and error, according to Live Science. The announcement builds on a moment that made IRON briefly famous: last year, presenters cut the robot open on stage to prove a person was not hiding inside its fabric-covered frame.
Specs behind the fabric skin: height, weight and 82 degrees of freedom
IRON ships in different body shapes and can be configured to present as different genders, but every version stands roughly 5 feet 7 inches tall (173 centimeters) and weighs 143 pounds (65 kilograms). The machine has 82 degrees of freedom from head to toe, including 22 in each hand, and can move at up to 6.5 feet per second (about 2 meters per second), a pace comparable to a fast power walk. A built-in self-charging function lets the robot locate a designated charging cabin and plug itself in when its battery runs low.
Xpeng has not said how many units it intends to build in this first production run, and the company’s public comments so far emphasize capability claims over delivery numbers. Even so, moving IRON from a hand-built prototype to a repeatable manufacturing line is itself a significant step for the humanoid robotics industry, where the gap between a working demo and a machine that can be produced consistently at scale has proven difficult for multiple companies to close.
Three Turing chips and a “world model” borrowed from Xpeng’s cars
Rather than developing a separate AI system just for the robot, Xpeng says it deployed the same underlying “world model” that already runs its AI-enabled electric vehicles and flying car prototypes. Each IRON unit carries three of the company’s Turing AI chips, delivering a combined 2,250 trillion operations per second of processing power. That is enough, Xpeng says, to run a model on par with Meta’s Llama 3.2 entirely on the robot itself, without an internet connection, compared with roughly 50 TOPS available in most laptops capable of simple on-device AI tasks like real-time webcam filters.
Combining vision, language and step-by-step reasoning
Xpeng has not disclosed the full architecture of IRON’s proprietary AI system, but says three frameworks work together to let it operate autonomously. A vision-language-action (VLA) component links visual perception, natural-language understanding and motor control; a vision-language model (VLM) pairs vision encoders with text processing; and a vision-language-task, or thought, model gives the robot the ability to reason through an assigned task in steps rather than acting on it directly. Researchers building humanoid robots generally cite the unpredictability of real-world environments as the core obstacle to reliable autonomy, which is why companies are increasingly investing in world models and similar planning layers instead of narrower task-specific programs. Robots that rely only on pre-scripted responses tend to work well in tightly controlled settings, such as a factory floor with fixed lighting and clear paths, but struggle the moment an object is out of place or a surface is uneven, which is precisely the kind of variability a household or a retail floor introduces.
Xpeng’s decision to reuse the same underlying world model across cars, flying car prototypes and IRON reflects a broader industry bet: that a system trained on enough varied physical data, regardless of the type of machine it originally came from, generalizes better than one built narrowly for a single robot form factor.
A cautious rollout: showrooms first, homes later
Despite the domestic ambitions attached to IRON, Xpeng plans to install the robots first in its own showrooms and technology campuses so it can keep validating use cases before wider release. External shipments to commercial customers are not expected until 2027, and those early buyers will likely deploy IRON for roles such as customer service or guided tours, primarily within China before expanding to other countries. Xpeng chairman and CEO He Xiaopeng said in a statement that the goal is to build “a new type of robot with full generalization capabilities that can truly become part of everyday life,” comparing the eventual spread of humanoid robots to the rapid adoption of smartphones in the mid-2000s.
Racing Figure, AGIBOT and a delayed Tesla Optimus
IRON enters a mass-production field that already includes a handful of rivals. American robotics company Figure AI says it has already ramped its own Figure 03 line to one robot per hour, a 24-fold increase in under 120 days that has pushed deliveries past 350 units, and Chinese firm AGIBOT says it had reached 15,000 units by June. Tesla’s Optimus, by contrast, remains in factory assembly and has had its release pushed back four times since the project began in 2022, most recently last year, a gap Xpeng representatives were reportedly careful to note in briefing materials shared with journalists. Whether IRON’s claimed advantages in shared AI architecture and onboard compute translate into a durable lead will depend on how quickly Xpeng can move beyond controlled showroom settings and into the messier environments of real homes and workplaces.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
