Shortly after launching its new humanoid production line and showcasing the first "Iron" unit walking autonomously off the assembly line, XPENG decided to reveal another side of the story: not just the factory itself, but the six years of errors, falls, and engineering decisions that brought the robot to this point. In a video lasting approximately six minutes, CEO He Xiaopeng presented a rare retrospective on Iron's development. Rather than focusing solely on current results, Xiaopeng recalled moments when the project was met with skepticism or even public ridicule.

When an early version of Iron appeared walking with human-like movements, accusations arose that a human might be hiding inside the machine; XPENG’s reaction was unusual—during a presentation, the company literally opened up part of the robot's leg in front of the audience to reveal the mechanical components beneath the outer shell.

Another incident occurred when a version of Iron fell during a demonstration at a shopping mall; footage circulated on Chinese social media, and the robot became a target of mockery. According to Xiaopeng, developing robots means learning to live with such failures. These experiences also helped shape the company's strategy: XPENG does not intend to focus its efforts on demonstrations designed primarily to produce impressive videos. While Xiaopeng acknowledges the progress made with humanoids capable of running, racing, or performing choreographed routines, he maintains that this is not the company's primary objective.

The goal is to develop a general-purpose humanoid capable of autonomous operation. In this context, Xpeng’s strategy is summed up by the phrase "the hardest path is never crowded," and part of that challenge lies in the intelligence controlling the robots. XPENG is leveraging technologies and infrastructure originally developed for its automotive systems; just as an autonomous vehicle must observe a constantly changing environment, identify objects, anticipate risks, and make real-time decisions, a humanoid faces a similar challenge but must translate those decisions into full-body movements. Consequently, Xpeng is steering the "Lion" project toward models that combine vision, language, and action. Rather than relying solely on pre-programmed sequences or remote human operators, the company aims for the system to interpret the environment and determine its own actions.

The video illustrates this approach with brief examples. In one instance, the system receives an image of a setting and is asked about the presence of a green sofa. Since the object is not in the image, the system correctly identifies its absence—a simple test designed to verify whether the system can interpret what it is actually seeing, rather than simply responding based on expected associations.

In another demonstration, the robot interacts with a person and concludes the conversation with a hand gesture. XPENG also aims to develop persistent memory, enabling the system to recognize different individuals and retain information from past interactions. The objective is to integrate visual perception, language, and physical control without relying on joysticks or covert teleoperation.

However, Xiaopeng also acknowledges that putting Iron into production does not mean development has ended; according to him, the first commercial version may not be perfect, but the robots rolling off the assembly line are capable of doing in the real world what they were developed to do: work autonomously.