Robotics will not get its “GPT moment” by following LLMs, Agibot chief scientist says
Luo Jianlan, chief scientist at Agibot and associate professor at the Shanghai Innovation Institute, argues that robotics will not achieve its "GPT moment" by simply mimicking the scaling laws of large language models (LLMs). He contends that while LLMs benefit from predictable statistical relationships between pretraining loss and model capability, embodied intelligence faces the complexities of the physical world, where offline loss reduction does not guarantee real-world deployment success. Luo emphasizes the need for a "deployment loop" where real-world interaction data continuously feeds back into model improvement, rather than solely focusing on data accumulation or parameter scaling. His work in China focuses on building this scalable loop through initiatives like SOP for online post-training, LWD for continuous learning during deployment, and the Tau-0-WM world model for action-conditioned physical prediction.
This perspective from Agibot's chief scientist highlights a critical divergence in the development paths of AI in China, specifically between large language models and embodied intelligence. While China has shown immense capability in scaling LLMs and leveraging vast datasets, Luo Jianlan's argument suggests that a direct replication of this strategy for robotics will be insufficient. This implies a need for Chinese companies to innovate beyond brute-force data accumulation, focusing instead on system-level integration and real-world deployment loops to drive robotic intelligence.
The emphasis on a "deployment loop" and the challenges in acquiring high-quality, real robot interaction data underscore a significant bottleneck for the embodied intelligence sector in Asia. China's strengths in manufacturing, supply chain, and engineering talent could be pivotal in establishing the necessary hardware infrastructure and deployment scale. However, the success will hinge on effectively integrating these elements with advanced AI models to create a self-reinforcing cycle of data collection, model training, and real-world performance improvement. This strategic shift could define the next wave of AI innovation in the region, moving beyond purely digital domains into tangible physical applications.
