Wang Xingxing on the Biggest Bottleneck of Physical AI: Alignment Between Models and the Physical World, Possibly Solvable in the Next 1 to 2 Years
Wang Xingxing, founder of Unitree Robotics, said at the 9th Hongqiao International Economic Forum that the core challenge of physical AI today is the alignment between models and the physical world, and that related problems are expected to be solved in the next 1 to 2 years. He pointed out that multimodal large models already possess strong understanding, planning, and navigation capabilities, but when running on real robots, sensor measurements at the input end are inconsistent with the true state, and action commands at the output end do not match actual motion, with errors being particularly pronounced in contact-based tasks such as grasping and assembly. IDC data shows that in the first half of 2026, global humanoid robot shipments approached 25,000 units, a year-on-year increase of 432.1%, with shipments in the Chinese market exceeding 19,000 units, accounting for about 77.9% of the global total.
Body · AI translation
Wang Xingxing Discusses the Biggest Bottleneck of Physical AI, Says It May Be Resolved in the Next 1 to 2 Years
Phoenix Tech Download the App for Exclusive Early Access November 05, 2026 18:22:09 From Beijing
Phoenix Tech News (Editor/Song Shumei) On November 5, at the "Humanoid Robots and Embodied Intelligence Industry Innovation and Development Cooperation" sub-forum of the 9th Hongqiao International Economic Forum, Wang Xingxing, founder and chairman of Unitree Robotics, said that the core challenge currently facing physical AI is the alignment between models and the physical world, and that the related issues are expected to be resolved in the next 1 to 2 years.
Wang Xingxing said that at this stage, multimodal large models already possess relatively strong understanding, planning, and navigation capabilities, but when running on real robots, deviations still occur between input and output. On the input side, there is the problem that sensor measurements are inconsistent with the true state; on the output side, there may be cases where the model's action commands do not match the robot's actual movement.
These errors are particularly evident in contact-based tasks such as grasping and assembly. Even if a robot can identify a target and plan an action, errors in the "last few centimeters" or even the "last few millimeters" may cause the task to fail, thereby affecting the model's generalization ability and the success rate of practical applications.
Previously, Wang Xingxing said that physical AI may see a substantive breakthrough in the next 1 to 2 years or 2 to 3 years, and Unitree Robotics is also conducting preliminary research on self-evolution models for physical AI robots, hoping to reduce the dependence of robot skill development on manual programming and repeated debugging through models autonomously collecting data, training, and iterating.
According to IDC data, in the first half of 2026, global humanoid robot shipments were close to 25,000 units, a year-on-year increase of 432.1%; shipments in the Chinese market exceeded 19,000 units, accounting for about 77.9% of the global total. As the scale of hardware shipments expands, the ability to align models with the real world is becoming a key link in the transition of robots from demonstrations to large-scale applications.