A Chinese humanoid startup flips 'distillation' claim on OpenAI as it releases a new robotics model
Zhu Mingxuan, who led the model's development, said she left U.S. humanoid company Figure last year, where she had also worked on models for helping humanoid robots mimic human actions.
Key takeaways
- The most recent developments show a rapid expansion of both civilian and military humanoid‑robot programs. In early September, Chinese electric‑vehicle maker XPeng announced that its IRON humanoid robot has moved from prototype to full‑scale manufacturing, walking off the production line on its own and slated to begin commercial shipments next year. The robot boasts 82 degrees of freedom, can move at roughly two metres per second and is being built in an automated line that is more than 80 % robot‑controlled.
- At the same time, Reuters reported that China’s People’s Liberation Army is actively converting advanced research into battlefield tools. After a high‑profile World Humanoid Robot Games in August, PLA officials called for faster transfer of “cutting‑edge technologies” into military training, and procurement documents from 2025‑2026 show tenders for embodied humanoid perception and dexterous‑operation systems as well as a $300,000 budget for data‑collection platforms to teach robots to navigate terrain and conduct tasks such as identity checks, discipline inspections and security patrols.
- American researchers have also pushed the performance envelope. A new AI framework called BeyondMimic, demonstrated by UC Berkeley and Stanford, enabled humanoid robots to sprint, spin‑kick and combine dozens of human‑like motions without task‑specific retraining, a breakthrough published in Science Robotics in September.
- Market analysts are projecting massive growth. Goldman Sachs now expects 6.5 million humanoid robots to be shipped globally by 2035, while companies such as Tesla, Teradyne and Intuitive Surgical are scaling production lines for service‑oriented robots.
- Elsewhere, Korea’s national AI project unveiled an upgraded KAIROS humanoid at the Global Forum on Mechanical Engineering, showing enhanced walking, hand‑shaking and culturally inspired dance motions. These parallel civilian, research and defense advances illustrate how quickly humanoid robots are moving from laboratory demos to mass production and potential deployment on the battlefield.
Zhu Mingxuan, who led the model's development, said she left U.S. humanoid company Figure last year, where she had also worked on models for helping humanoid robots mimic human actions. She told CNBC earlier this week that Aether reduced training time by two-thirds, and that she planned to open-source parts of the model, such as those relating to touch, but not portions related to energy use.
Getting humanoids and AI models to perform as intelligently as humans across a variety of tasks has remained a challenge, amid a broader tech race between U.S. and Chinese companies. In this article
The CEO of an Ant-backed humanoid robotics startup has published a letter to OpenAI in Chinese that raises questions about technical and design similarities between the two companies' recent models.
"People often say major tech companies have intelligence networks monitoring the whole internet, this time I believe it, this is a direct distillation of us without any modifications," Guo Renjie, CEO of Suzhou-based JoyIn, said in a public statement Thursday, according to a CNBC translation. CNBC was unable to independently verify the claims. Some concepts are existing parts of AI research more broadly, although companies connect and implement them differently. OpenAI did not immediately respond to a request for comment.
U.S.-based Anthropic has repeatedly flagged unauthorized "distillation" of its models by Chinese companies to improve their own AI capabilities. On Tuesday, a U.S. cybersecurity agency said six Chinese companies, including DeepSeek and Alibaba, distilled models from Anthropic, Google and OpenAI.
JoyIn's Aether model, which Guo said he decided to publicly announce Thursday, claims its perceptive, rather than text-based, approach to robotic control enables humanoids to complete tasks with a 90% success rate on first attempt.
New robots, funding, and signals. Every Thursday.
Unsubscribe anytime.
Mentioned in this article