Humanoid robots have emerged as a core track in the transformation of the global tech industry. However, constrained by technological bottlenecks, they have not yet truly entered mass household scenarios. Over the next few years, the iteration of humanoid robots will accelerate, and large model-driven robot self-evolution is poised to become a key evolutionary trend.
On August 20, Wang Xingxing of Unitree Robotics delivered a keynote speech titled "From Exhibits to Products: The Next Decade of the Humanoid Robot Industry" at the 2026 World Robot Conference. He pointed out that the insufficient generalization capability of embodied AI is the primary bottleneck hindering its mass adoption. Nevertheless, the humanoid robot market is expected to hit an explosive growth point in as little as 2 to 3 years.
Insufficient Generalization Capability, but an Explosion Point Will Arrive in the Next 2 to 3 Years
Currently, general-purpose humanoid and semi-humanoid robots face challenges in large-scale scenario deployment. Wang Xingxing believes that the insufficient generalization capability of embodied AI remains the primary bottleneck preventing robots worldwide from entering households and integrating into daily life.
In his speech, Wang Xingxing noted that currently, in fixed scenarios, global AI models can achieve a task success rate of nearly 100% after sufficient data collection and specialized training. However, simply changing the manipulated objects or slightly adjusting the surrounding environment causes the robot's task execution success rate to plummet, rendering it incapable of adapting to the complex, dynamic, and unfamiliar scenarios encountered in real life.
"Currently, global embodied AI models are primarily divided into two mainstream paradigms: VLA models and world models, with world models being the key focus of the industry in 2026," Wang Xingxing stated.
Wang Xingxing introduced that AI model R&D is currently the core business direction receiving the largest financial and human resource investments at Unitree Robotics. Unitree initiated the R&D of video-generation-based world models as early as early 2024. However, the R&D outcomes by the end of 2024 fell short of expectations, leading to a temporary suspension of the related work. In 2025, Unitree Robotics ramped up its investments once again, continuing to deeply engage in the video-generation-based world model track.
In Wang Xingxing's view, humanoid robots are on the verge of reaching an explosive growth point. Depending on the pace of iteration, this milestone will be reached in as little as 2 to 3 years, or 5 to 10 years at a slower pace.
He believes that when a robot can adapt to various unfamiliar household scenarios and independently complete approximately 80% of daily tasks based on voice and language commands, it signifies that the embodied AI industry has reached a critical inflection point for explosive growth. This also serves as the core milestone for robots transitioning from exhibits to mass-produced products.
Currently, general language models and multimodal models operate on pure digital encoding, with inputs and outputs confined to a fixed vector space, resulting in virtually zero deviation or loss. Wang Xingxing emphasized that robots are physical entities; every perceptual input and action output generates deviation and loss. This is the core reason why the generalization capability and real-world success rate of robot models are significantly lower than those of digital large models.
Addressing the long-standing technological bottlenecks in the industry, Wang Xingxing stated that the most significant issue currently is the insufficient alignment between the inputs and outputs of AI models and the actual physical world of robots. Existing models can accurately execute overall task logic, and the general execution processes—such as robot movement, assembly, and material handling—are largely error-free. However, fixed deviations occur during fine operations at the centimeter and millimeter scales. In practical operations, while the robot's visual system may nearly complete a grasping action, it fails to correct minute physical errors or compensate for the lack of tactile feedback, ultimately resulting in task failure and a drastic plunge in the success rate. This remains a common challenge across the global embodied AI industry.
Large Models Drive Robot Self-Evolution to Enhance Iteration Efficiency
"For a long time, the AI development paradigm in the robotics field has been relatively traditional. The approach relying on manual development and debugging suffers from low efficiency, and the industry lacks practical, real-world applications of AI technologies," Wang Xingxing noted in his speech.
To address this, Unitree Robotics leverages cutting-edge, top-tier AI large models. By defining custom R&D rules, experience constraints, and toolchains, the company enables robot agents to autonomously search for the latest global academic papers, high-quality research findings, and various open-source solutions, automatically generating dedicated robot control code.
Wang Xingxing introduced that during the performance validation phase, Unitree Robotics employs an evaluation model combining "autonomous robot scoring and manual assessment" to ensure the results better align with real-world deployment needs. Meanwhile, leveraging the Copilot agent, Unitree Robotics has automated the programming of deep reinforcement learning algorithms, simulation training, and real-machine testing, significantly boosting the efficiency of robot development and iteration.
Discussing the deployment of robot self-evolution, Wang Xingxing stated that this model offers significant industry advantages. On the one hand, it can upgrade synchronously alongside the continuous iteration of global foundational AI models, steadily enhancing the robot's self-evolution capabilities and forming a virtuous cycle of technological growth. On the other hand, it can activate multi-source data from real-world scenarios and human interactions, resolving the pain point of low data utilization in traditional manual development.
As the scale of real-machine deployment continues to expand, companies can accumulate massive amounts of test data and evaluation metrics, achieving large-scale data precipitation. This allows robots to continuously acquire skills, preventing the loss of R&D outcomes, thereby improving both the efficiency and speed of robot development and evolution.
Currently, with the comprehensive boom in AI technologies and the continuous surge in global industry attention, the evolutionary speed of robotics has accelerated significantly. The pace of industrial development over the next decade may well exceed market expectations.
-END-
Editor-in-Charge / Layout: Sun Tian
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng