Unitree founder says embodied AI's 'ChatGPT moment' hinges on closing millimeter-level gap

Wang Xingxing, founder of Unitree Robotics, delivered a keynote speech titled "From Mechanics to Intelligence: The Evolution of Embodied Future" at the 5th Global Digital Trade Expo in Hangzhou on September 24, sharing the company's latest robot technology progress and his views on the industry.

Wang said that over recent months, Unitree's robot foundation model has made continuous progress and can now generate and execute robot actions in real time based on human voice commands. Due to the real-time computation involved in speech understanding and action generation, there is still a slight latency during execution. The robots are now capable of performing long-sequence continuous tasks and resisting interference, though operational efficiency still needs improvement. He hopes that by the end of 2026 or 2027, anyone will be able to give a robot an instruction and have it generate the corresponding action, enabling large-scale real-world deployment.

Looking back at the development of the AI industry, Wang noted that ChatGPT reshaped public perception of artificial intelligence. He predicted that embodied intelligence will also see its own "ChatGPT moment." When robots can complete 80 percent of tasks in 80 percent of unfamiliar scenarios through voice-driven embodied capabilities, the industry will reach a tipping point, which he expects could arrive within the next few years.

The biggest challenge, in his view, is the insufficient precision matching between AI model inputs and outputs and the real physical world, where robots still have errors of a few millimeters. Whoever solves this problem will solve a fundamental bottleneck of the robotics industry, Wang said.

He also pointed out that the current global development model for robotics is still relatively primitive, with heavy human involvement in code generation, training deployment and real-robot testing, leading to scarce real-world data. Wang proposed replacing the traditional manual process with an autonomous AI loop, where AI generates code, trains models, deploys to robots, conducts tests and iterates automatically. Unitree could then scale from testing dozens of robots per day to hundreds or even thousands.