Unitree's Wang Xingxing: 2027 is very likely to be the critical turning point for the robotics industry
Wang Xingxing, founder of Unitree Technology, judged at the 2026 World Internet Conference Wuzhen Summit that 2027 has a very high probability of becoming the critical turning point for the robotics industry or the development of robot models. He pointed out that the current robot data utilization rate is low, with large amounts of data being wasted, and that AI models need to be allowed to write code, train in simulated environments, and provide feedback iteration after deployment testing. He said that the "GPT moment" of embodied intelligence is when robots complete about 80% of tasks in about 80% of unfamiliar scenarios through voice commands; the world's top models are already very close, but physical-world contact and alignment are still short by "the last few millimeters."
Body · AI translation
On November 4, at the "Internet Entrepreneurs Forum" of the 2026 World Internet Conference Wuzhen Summit, Wang Xingxing, founder of Unitree Robotics, spoke about technological breakthroughs in the robotics industry, saying that at present, including at Unitree, the utilization rate of robot data remains low, and a large amount of data is not fully utilized. He believes that as physical AI models achieve self-evolution, robots are expected to improve data utilization efficiency and accelerate technological iteration.
"For us, manufacturing and deploying a thousand or ten thousand robots to test and collect data is, in a sense, a very low-cost and relatively easy thing to do," Wang Xingxing said. At present, the company's robot data utilization rate is somewhat low, and "a large amount of data is being wasted; AI models have not been put to good use."
But this is still relatively rare in the robotics field. He said the logic is to have AI models write code and train in a simulation environment, then deploy them to robots for testing, and after testing is completed, humans or robots can assign scores. These results are then fed back to the AI model so that it can revise the program and train again.
In this process, what value do humans mainly provide?
"Up to now, I think the greatest value of humans lies in 'scoring'—evaluating whether the AI model has done a good job, or setting some goals for AI to achieve," Wang Xingxing said. He also mentioned that as AI models become stronger, they will also learn human evaluation metrics and preferences. At present, in the video model field, this effect is relatively intuitive. Wang Xingxing believes that as models undergo self-evolution and iteration, their data utilization rate will also be much higher.
As for when the robotics industry will achieve a breakthrough, Wang Xingxing gave a relatively optimistic judgment. He said that 2027 has a very high probability of becoming a critical point for the development of the robotics industry or robot models.
Wang Xingxing mentioned that the "GPT moment" of embodied intelligence is when robots can complete about 80% of tasks through voice commands in about 80% of unfamiliar scenarios. This time, he said that the world's top models are already very close to this goal, but key technical obstacles remain.
He called this obstacle "the problem of the last few millimeters."
He explained that current top multimodal models already have strong spatial understanding and task planning capabilities. After sufficient training, robots performing a single task at a fixed workstation can achieve a success rate close to 100%, but operational efficiency still needs improvement. For tasks such as grasping and carrying, models can already achieve relatively high success rates, but in scenarios requiring fine operations, such as phone disassembly and assembly or threading a needle, success rates remain relatively low.
"AI models do not do a particularly good job in the part involving contact with the physical world—that is, the alignment between the input and output of AI models and the real world is still a bit off," Wang Xingxing said. He believes that robots still lack sufficiently accurate understanding of information such as contact and touch with the physical world, and this is precisely the key to completing complex operations. He expects that such problems are expected to be solved in the next one to three years.
However, these judgments by Wang Xingxing are still predictions about technological progress. A high task success rate at a fixed workstation does not mean that robots already possess general capabilities for working across scenarios, and the mechanism of model self-evolution still needs further verification.
Beyond expectations for technological breakthroughs, Unitree's performance in the capital market has also attracted attention. On November 4, Unitree Robotics (688836.SH) shares rose by more than 8% at one point during intraday trading, surpassing Yuanjie Technology in share price and ranking fourth among A-share stock prices.
For this company, which has already achieved large-scale robot manufacturing, competition in the next stage lies not only in how many robots it can produce, but also in whether it can enable models to make full use of data, so that robots can move from completing specific actions to autonomously completing tasks in more unfamiliar scenarios.