Workflow
北京人形机器人创新中心公布多项成果 加速具身智能迈向产业实用

Core Insights - The Beijing Humanoid Robot Innovation Center unveiled four core achievements in embodied intelligence at the 2025 World Robot Conference, marking a shift from technological breakthroughs to industrial practicality [1][2] - The "Embodied World Model System" serves as the core brain for robots, enabling them to understand reality and predict changes, featuring a 72B multimodal large model and a neural network-driven world simulator [1] - The "Cross-ontology VLA Model" allows for flexible adaptation of robots across multiple ontologies and scenarios, demonstrating robust generalization capabilities and significant potential for rapid skill acquisition [1][2] Industry Challenges - Current VLA models face challenges such as action prediction difficulties, significant differences in robot ontology, data incompatibility, and weak task generalization [2] - The direct mapping of image pixel space to robot operation space presents optimization challenges and prediction inaccuracies, complicating the application of models across different robot configurations [2] Data Collection Initiative - The "Thousand Robot Real Scene Data Collection Plan" aims to enhance core capabilities by deploying robots in real industrial environments, collecting multimodal interaction data during actual operational tasks [2] - This initiative contrasts with laboratory settings, as real-world environments introduce variability and unexpected conditions, providing richer training material for embodied intelligence models [2] Future Outlook - Industry experts agree that industrialization is essential for the development of embodied intelligence, emphasizing the need for a standardized and replicable technological leap to achieve scalable implementation [2]