中金发布研报称，机器人大模型是破解传统机器人控制瓶颈、迈向通用具身智能的关键路径。当前行业 主要基于大语言模型、自动驾驶大模型及多模态大模型探索的发展方向，产业重心已转向"小脑+大 脑"系统研发，而不同企业在研发与商业化路径上存在差异。该行认为只有极少部分具备全栈技术能 力、资源整合优势与长期主义战略的企业，未来将通过收敛技术路径，最终定义"具身智能"的核心标 准。 中金主要观点如下： 机器人大模型助力通用具身智能发展 传统机器人在任务、场景和数据方面存在较强的专一性，泛化能力较弱，难以应对复杂环境，更偏 向"人形机器"的属性。相比人类学习，机器人在集体学习效率上具备优势。目前行业已形成共识，即机 器人大模型可通过融合视觉、触觉等多模态信息，弥补机器人在"物理常识"方面的不足，是推动产业向 通用具身智能迈进的重要路径之一。 商业化路径选择与企业能力边界待明确 商业化进程中，"硬件优先"(由车企、机器人企业主导)与"模型优先"(由AI企业主导)两种路径各有特点 与优势。受场景复杂度、技术门槛以及商业回报周期等因素影响，多数企业可能会聚焦于特定垂直领 域，实现该场景下的"通用/柔性"应用;该行认为，仅有少数具备全 ...

Core Viewpoint - The development of large models for robotics is seen as a key pathway to overcoming traditional control bottlenecks and advancing towards general embodied intelligence [2][4][18]. Group 1: Importance of Large Models in Robotics - Large models can address the fundamental issue of robots lacking physical "common sense" by integrating multimodal information such as vision and touch [4][18]. - The industry consensus is shifting towards the development of "small brain + big brain" systems, indicating a focus on foundational capabilities for robots to be applied in various scenarios like smart manufacturing and home services [18][36]. - The transition from humanoid robots to systems that leverage large models reflects a response to national strategies and societal needs, particularly in addressing labor shortages in service industries [18][36]. Group 2: Limitations of Existing Models - Current mature models, such as large language models, have limitations in directly solving physical operation problems for robots and often exhibit "hallucination" phenomena [4][24]. - While large language models excel in natural language processing, they cannot fully empower robots due to their inability to understand physical world causality, which is crucial for executing tasks in real environments [24][26]. - The challenges faced by robots are more complex than those in autonomous driving, requiring greater generalization and adaptability to diverse and unstructured environments [4][24]. Group 3: Commercialization Pathways - Two primary commercialization pathways are identified: "hardware-first" led by automotive and robotics companies, and "model-first" led by AI companies, each with distinct advantages [5][40]. - Most companies are likely to focus on specific vertical applications, achieving "general/flexible" capabilities, while only a few with full-stack capabilities may define the standards for "embodied intelligence" [5][40][43]. - The market is experiencing a significant increase in investment, with a reported 80% growth in financing events in the first half of 2025 compared to the same period in 2024, indicating heightened interest in the robotics sector [36]. Group 4: Future Trends and Challenges - The robotics industry is expected to evolve towards a model of specialized division of labor, moving away from the current "full-chain self-research" approach [46]. - The gap between market expectations and actual robotic capabilities continues to widen, with increasing demands for robots to perform complex tasks beyond simple automation [37][38]. - The integration of multimodal capabilities is essential for enhancing robots' perception and task execution, as traditional methods struggle to provide comprehensive environmental understanding [27][29].

Group 1 - The 2025 World Robot Conference showcased over 200 domestic and international robot companies, with more than 1,500 exhibits and over 100 new product launches, highlighting the rapid growth and innovation in the robotics industry [1] - In 2024, China's industrial robot market is projected to sell 302,000 units, maintaining its position as the largest industrial robot market globally for 12 consecutive years [1] - China accounted for two-thirds of global robot patent applications in 2024, indicating a strong focus on innovation within the industry [1] Group 2 - The complete industrial system and diverse application scenarios are driving rapid iterations in China's robotics sector, particularly in humanoid robots, following the Ministry of Industry and Information Technology's guidance on innovation [1] - Domestic companies have achieved nearly 100% self-research and development in core components like joint modules and sensors, providing a competitive edge over international counterparts [1] - The rise of large models in recent years has further supported the robotics industry, with humanoid robots aiming for "general intelligence" and emphasizing a complete chain of perception, decision-making, and execution [1][2] Group 3 - The rapid evolution of large models in robotics is attributed to breakthroughs in both architecture and data, with models now capable of understanding tasks and generating actions [2] - A structured data support system has been established, combining internet, simulation, and real machine action data to enhance model capabilities in real-world applications [2] - Future developments in embodied large models are expected to focus on modal expansion, reasoning mechanisms, and data composition, potentially incorporating additional sensory channels [2] Group 4 - At the robot conference, over 50 complete machine companies presented hundreds of humanoid robot products, expanding application scenarios [3] - The evolution of robots is anticipated to accelerate, with general-purpose robots expected to gradually enter the consumer market within the next 3 to 5 years [3] - Two key value considerations for robots in practical applications are the data value of tasks and the economic value provided to customers by automating tedious and low-value labor tasks [3]