Wang Xingxing: Humanoid Robots' Biggest Shortcoming Is Alignment, Errors in Last Few Centimeters

On September 9, Wang Xingxing, speaking at the 2026 Private Economy Innovation Development Conference, addressed the next decade of the humanoid robotics industry. He judged that the industry's biggest current shortcoming is alignment.

Wang noted that today's humanoid robots can dance, flip, and complete many commands, but in many scenarios they cannot perform the same tasks after the environment changes. The alignment between AI model inputs and outputs and real robots is insufficient, leading to errors when robots actually work. This misalignment means that while robots can perform impressive motor skills, they lack the reliability needed for everyday tasks.

Wang Xingxing explained that the error appears in the last few centimeters or even millimeters. This tiny gap blocks improvements in generalization and the robot's success rate in entering homes. The issue is a concrete engineering problem in embodied intelligence. Actions planned in simulation or training data deviate when applied to real robot joints, sensors, and end effectors due to calibration errors, mechanical gaps, sensor noise, and material deformation.

These deviations may not be obvious during large-scale movements, but in grasping, placing, and insertion tasks that require precise positioning, a few millimeters can cause failure. This is why robots can run and jump but still struggle with fine manipulation. It is also the key obstacle preventing humanoid robots from moving from demonstrations to practical use. The discrepancy between planned and actual movements accumulates, especially in tasks that require precision.

Wang described a scenario: a person brings a robot that has never entered a home and casually asks it to tidy a room. Without prior training, the robot can organize the messy room. He believes the robotics industry may then see explosive growth. This scenario sets a much higher bar than laboratory tasks. It requires the robot to enter a completely unfamiliar environment, without prior mapping or training data, and rely on general capabilities to understand messy semantics, decide where items belong, plan a tidying sequence, and complete grasping and placement. Without alignment, the robot cannot reliably perform these steps.

Achieving alignment is difficult because the real world introduces uncertainties that simulation cannot fully capture. Every robot has unique physical characteristics, and wear and tear over time can change its behavior. As a result, a model that works perfectly in one robot may fail in another.

Wang had said on August 20 at the World Robot Conference that the biggest bottleneck is insufficient alignment between AI model inputs and outputs, but he believes this will inevitably be solved in the coming years. This earlier statement underscores his consistent focus on alignment as a core issue.

Unitree Technology listed on the STAR Market on August 19, dubbed the first humanoid robot stock. After an initial surge, its share price fell below 500 yuan on September 14, down 57.27% from its peak, with total market value shrinking to 190.1 billion yuan. At this juncture, Wang's discussion of alignment pulled the conversation from capital market valuation fluctuations back to fundamental industry issues. The stock market reaction reflects broader volatility in the humanoid robotics sector, but Wang's comments redirect attention to technical challenges.

He gave a timeline: the problem can be solved in a few years, but the speed depends on how well joint optimization between models and hardware can be achieved. Unitree allocated 48% of its IPO proceeds to model research and development. This investment direction is consistent with his alignment judgment and indicates that Unitree views technological breakthroughs as the core of future competitiveness. Wang's emphasis on alignment suggests that progress in humanoid robotics will depend not only on better AI models but also on tighter integration with hardware.