On September 10th, Chinese embodied intelligence company Lexiang Technology publicly accused OpenAI of "pixel-level plagiarism," a claim that went unanswered by the latter but drew skepticism from public opinion and industry insiders, who questioned it as "hype-chasing" and "demo cheating." In response, the company launched a "self-vindication" livestream on September 16th. At 5:30 PM, the broadcast commenced.
Lexiang Technology described the event as the "world's first real-world outdoor livestream of robots in an open environment," a departure from typical demos that can be run repeatedly, adjusted, or even spliced together. The company's plan was to showcase robots accepting random customer orders, serving food, and interacting in a dining setting over the course of an hour. Using a barbecue stall as the backdrop, Lexiang deployed its humanoid robots, nicknamed "Er Ye" and "Wu Ye," to complete service tasks for three groups of customers.
"Authenticity" was the central theme of the entire broadcast. Mid-stream, Lexiang's CEO Guo Renjie personally guided the camera through the back kitchen, flipped the lens to film the crew, and pulled back a backdrop curtain in an attempt to prove the robots' operations were fully autonomous and not remotely controlled. Yet, the question remains: does this livestream sufficiently validate the substance of Lexiang's technology and products? Many questions were left unanswered.
After issuing three challenges to OpenAI, Lexiang Technology hastily organized this livestream to defend itself. At 5:30 PM, the feed started, and everything appeared rushed and improvised: audio intermittently picked up chatter among staff; Guo Renjie, serving as his own host, occasionally issued impromptu commands like adjusting camera positions or saying "do it again." Staff in blue t-shirts, two "customers" per group acting as robot patrons, and the robots' positioning all seemed unrefined, repeatedly blocking the camera view.
Guo Renjie did not shy away from the lack of preparation. After the opening, he revealed: "We only finalized the deal with this barbecue shop owner on Sunday (September 13th), renting the venue for tens of thousands of yuan." The chaos, ironically, added to the sense of realism, which was likely the effect Lexiang wanted to convey. After the robots completed two rounds of service, Guo stated: "We only arrived on-site at 4 PM today. Now, let's show everyone whether 'remote control' actually exists." He then guided the camera through the kitchen, flipped the lens to film the crew, and lifted the backdrop cloth to demonstrate the absence of any remote operators.
This hastily convened livestream was widely interpreted by industry observers as an escalation in Lexiang's public relations offensive. On September 10th, Guo Renjie published an article on the "Lexiang Technology Joyln" WeChat official account. It first presented a 10-minute demo video of two robots collaboratively handling household chores, claiming the development of Aether, a cross-embodiment, general-purpose, self-evolving, multi-robot collaborative embodied intelligence model. In the same piece, Guo directly named OpenAI Chief Scientist Jakub Pachocki, alleging that his widely noted September 6th essay, "An Alien Mind," had "pixel-by-pixel borrowed" the research findings of the Aether model. Guo declared plans to initiate legal proceedings against OpenAI over the matter.
A startup founded in December 2024, Lexiang Technology claimed to have achieved "intelligent emergence" in robotics and leveled plagiarism accusations at a leader of the global large model wave—a move that quickly drew attention. However, the buzz was accompanied by doubts about Lexiang leveraging OpenAI's fame, marketing theatrics, demo fraud, and insufficient technical disclosure. At least on the surface, this livestream provided more information than the previously released demo.
From the livestream's performance, the robots moved slowly and occasionally stalled, but they managed tasks like greeting customers, taking orders, fetching ingredients, grilling food, seasoning, handing over dishes, and even giving high-fives and handshakes. Notably, in specific tasks like flipping food and adding seasoning, the robots initially underperformed but showed better results on a second attempt—described by on-site staff as "it self-optimizes, learns, and re-plans." Viewers could observe the robots' continuous operation, including failures, pauses, and the staff's on-the-fly reactions.
Before the broadcast, Lexiang did not disclose the specific venue or start time in advance, nor did it offer media or third-party on-site verification. Perhaps consequently, the live audience was small—when the camera panned over the scene, around 20 spectators were visible. In contrast, the comment section was livelier: as the robots performed, phrases like "the authenticity checks out" and "that serving scene feels truly autonomous and real" scrolled by.
An hour-long livestream isn't enough to answer every question. The event brought Lexiang more attention, but it clearly hasn't dispelled all skepticism. According to insights from industry insiders, many feel that the "realness" of the livestream does not equate to the "realness" of Lexiang's technical capabilities. Qiu Lai, an employee at a Beijing-based embodied intelligence firm, commented: "First, on pure motor performance—delivering items, flipping food, gait—there are definitely better robots out there. Second, Lexiang's article claims Aether handles fully open tasks and self-evolves. You can't confirm or refute that with just one livestream." He added, "I watched five minutes and felt it was a waste of time."
Regarding the robot "Wu Ye's" multiple attempts at seasoning, where the first try failed and the second succeeded, Qiu argued this doesn't prove "self-evolution." "There's another possibility: action execution has inherent randomness. Have you ever seen a robot dog flip once, fall, then stand on the second try?" In his view, whether the model has truly "evolved" or capabilities can transfer across embodiments requires longer-term validation and observation. Another embodied intelligence practitioner echoed these doubts, noting that based on the footage, "you can't discern the robots' visual or haptic feedback."
Beyond the livestream, Lexiang Technology and Aether face further scrutiny. In the article accusing OpenAI, Lexiang summarized Aether's training with "200, 0, 4"—approximately 200 hours of human video, zero real-world robot data, and a 4 billion parameter model. Lexiang also claimed Aether achieves over 90% zero-shot task success in unfamiliar environments. These claims raise numerous questions: What data modality was used for the 200 hours of training video? How precise was the annotation? Was the data collected in-house or from public datasets? What tasks were covered? To answer these questions, Lexiang must still address public standardized evaluation and allow independent third-party replication.
Before founding Lexiang Technology, Guo Renjie was best known as the Executive President of Dreame Technology's China division and a "marketing genius." In 2021, he gained fame with a viral video of a floor washer absorbing instant noodles, which quickly opened up the market—according to official sources, he grew Dreame's floor washer revenue from zero to millions of yuan in just 80 days. But in the embodied intelligence market, every viral move demands extended validation and more complete technical evidence. When a company claims to have achieved "intelligent emergence," the burden of proof becomes center stage.
Public information indicates that Lexiang Technology has laid out product lines including small humanoids, tracked robots, home collaborative robots, and full-size humanoid robots. Its high-performance small humanoid robot, OpenBridge, priced at 8,888 yuan, is already on sale at JD.com. As of this writing, Lexiang has not released a full technical report on the Aether model.
Comments