On September 23, a seminar on the high-quality development of embodied intelligence was held in Beijing, organized by the Digital Economy Research Center of the China Economic Information Service. Zou Zhensheng, Vice President of Engineering at Mianbi Intelligent Technology Co., Ltd., emphasized that embodied intelligence is not a solitary endeavor but requires ecosystem collaboration. He stated that the final mile for scaling embodied intelligence is not about impressive demonstrations but rather about reliability and ensuring economic feasibility for clients.
Mianbi Intelligent, established in 2022 from the technology transfer of Tsinghua University's Natural Language Processing Laboratory, focuses on the importance of different data ratio strategies at various training stages as a key to model evolution. Zou pointed out that human-first perspective data is currently the only supply capable of scaling to tens of millions of hours, making it a potential foundation for pre-training in embodied scaling laws, despite lacking action labels and ontology perception.
Zou identified three dimensions that determine the pace of embodied intelligence implementation: task definability, the cost of failure, and the ease of data feedback. He categorized application scenarios into four tiers, with automotive and logistics sorting at the forefront due to their high standardization and smooth data feedback. He noted that the core constraints on large-scale commercial use are not only model capabilities but also reliability and system engineering, stressing the need for economic viability for clients.
Editor's Note
The insights from Zou Zhensheng reflect a growing recognition in the robotics industry of the importance of reliability and economic feasibility in the deployment of embodied intelligence. As companies seek to integrate advanced AI technologies into practical applications, the collaborative ecosystem approach he advocates may be crucial for overcoming existing challenges and achieving scalable solutions.
Leave a comment