Summary
- Wang Xiaogang, Chairman of ACE Robotics, forecasts that embodied AI may reach its "ChatGPT moment" by late 2027.
- A lack of sufficient real-world training data is a significant hurdle, with ACE planning to gather tens of millions of training hours over the next two years.
- The Chinese startup intends to implement its technology in 1,000 stores within the year and eventually aims for an IPO.
While humanoid robots can walk, dance, and box, effectively performing useful tasks in unpredictable real-world settings remains a significant challenge for artificial intelligence.
According to a Reuters report, ACE Robotics Chairman Wang Xiaogang believes that advancements in AI models and the accumulation of real-world training data will soon enable robots to transition from mere demonstrations to practical applications.
"We anticipate reaching the 'ChatGPT moment' for embodied intelligence by the end of next year, fueled by world models and the capture of environmental data," Wang stated in the Reuters article.
Established in July 2025, ACE Robotics is a Chinese startup focused on developing AI models tailored for humanoid robots. Supported by Ant Group and SenseTime, the company raised over $100 million in the first half of 2026 and is considering an IPO "as soon as possible," as reported by Reuters.
Despite the rapid proliferation of large language models like ChatGPT and DeepSeek, robots continue to face challenges in executing a variety of tasks in unfamiliar situations, largely due to the scarcity of training data.
Embodied AI enables robots and other physical agents to sense their surroundings, reason about them, and translate decisions into actions using sensors and actuators. Concurrently, world models assist AI in understanding the dynamics of the physical environment by learning how objects and settings interact.
This capability allows robots to predict potential outcomes when they move, pick objects up, or engage with their environment before taking action.
Wang noted, "In recent years, the entire industry has amassed about 100,000 hours of data, which is far from adequate for training embodied foundation models."
Other research teams are exploring similar methodologies. In October, a team introduced HumanoidExo, a wearable exoskeleton designed to capture human movements for training humanoid robots.
Various companies are also crafting their own AI models for robotic applications. In January, Boston Dynamics unveiled the commercial version of its Atlas humanoid robot, citing AI advancements that bring it closer to real-world deployment. In June, Alibaba launched the Qwen-Robot Suite, a collection of AI models aimed at helping robots navigate, carry out physical tasks, and replicate real-world environments.