Wang Xingxing: Robot Combat Is Less Entertaining, and the Biggest Challenge for Embodied Intelligence Is Aligning Models with the Physical World

“Let me share something interesting: after a year, robot fighting has actually become less entertaining to watch.” Wang Xingxing, founder and chairman of Unitree Robotics, said this yesterday (the 5th) when he appeared at the “Cooperation on Innovative Development of the Humanoid Robot and Embodied Intelligence Industry” sub-forum of the 9th Hongqiao International Economic Forum. Wang Xingxing. Wang Xingxing played on site a video of Unitree robots fighting from two months ago. “A while ago, we achieved two robots fighting fully autonomously. They had nearly 20 kinds of movements, and in theory robots can be trained in hundreds or even thousands of movements, which can be combined arbitrarily.” Recall that in the first half of last year, Unitree Robotics launched the world’s first humanoid robot fighting competition. “But in just about a year, everyone will notice that the fights are less entertaining to watch than last year. Why not as good as last year? Actually, it is because last year the robots’ stability was not as good as it is now, so the sense of combat confrontation was stronger.” Wang Xingxing explained that this year, the robots’ stability has improved greatly; no matter how they are hit, they can stand firm, so scenes of falling down and collapsing naturally become fewer. Robot fighting. Wang Xingxing also talked about what he considers the biggest challenge currently facing robots or embodied intelligence. “The biggest challenge is not that multimodal models are weak in planning or understanding. In fact, multimodal models are already very strong in these areas.” He gave an example: not long ago, under “zero training,” GPT-6 directly controlled a Unitree robot to paint. “Changing the water for the brush, mixing paint, and so on were all directly driven by the model, and the robot’s work was very impressive.” Robot painting driven by a model. Wang Xingxing instead believes that the biggest current challenge for embodied intelligence lies in the poor alignment between the inputs and outputs of multimodal models and the physical world. “Right at the last few millimeters, there is still a relatively large deviation. If the world’s top multimodal model is used to directly control a robot to grasp, the success rate is almost 100%. But if it comes to more precise operations on components, the effect is much worse.” Wang Xingxing is confident that the alignment between the inputs and outputs of multimodal models and the physical world should be well solved within the next one to two years. “Embodied intelligence itself no longer has many technical difficulties, though its deployment and promotion still require solving a large number of engineering problems.”