“泛化智能”发展态势渐趋明朗 机器人技术路径出现分化

Core Insights - The 2025 World Robot Conference highlighted the trend of "generalized intelligence" in robotics, with VLA (Vision-Language-Action) models being a prominent topic across various exhibits [1][2] - The human-like robots showcased at the conference demonstrated advanced capabilities in dexterous operations, dynamic behaviors, and adaptability to complex environments, marking a shift from precision efficiency to generalized intelligence [1][2] Group 1: VLA Model Applications - Beijing Galaxy General Robotics introduced a humanoid robot capable of efficiently retrieving items from densely packed shelves, utilizing their self-developed Grocery VLA model for adaptability in diverse environments [2] - Beijing Star Motion Era Technology showcased humanoid robots collaborating in logistics sorting, powered by their end-to-end VLA model, enabling complex movements and continuous service evolution through a feedback loop of data collection and model training [2] Group 2: Differentiated Technical Paths - Leju Robotics presented their full-size humanoid robot "Kua Fu," capable of performing various industrial and commercial tasks, demonstrating the transition of humanoid robots from laboratory settings to practical applications [3] - The company employs a layered planning and model-based reinforcement learning approach, integrating multimodal embodied models for environmental perception and action generation [3][4] Group 3: Commercialization and Future Outlook - UBTECH Robotics emphasized the importance of rapid commercialization, predicting a significant growth phase for the humanoid robot industry around 2027, with expected annual deliveries reaching tens of thousands [5] - The industry is witnessing varied approaches to VLA and reinforcement learning technologies, with the architecture being a critical factor influencing training costs and efficiency [5]