Northeast Securities: The evolution of robot training towards the integration of virtual and real, and the AI-driven general training paradigm; physical AI has great potential

According to the Zhitong Finance APP, Northeast Securities has released a research report stating that the robot training industry is undergoing a transformation from industrial customization to a virtual-physical integration and AI-driven generalized training paradigm. Generative AI enables users to quickly generate new content based on various inputs; in addition, zero-shot and few-shot learning will become more prevalent, and the combination of generative models with large language models will promote zero-shot generalization of robot tasks. In terms of physical AI, robots need a deeper understanding of the physical world to cope with complex real-world scenarios, becoming partners that help humans solve complex problems practically. Related targets include Suochen Technology (688507.SH), SHARETRONIC (300857.SZ), and Qunke Technology.

The main points of Northeast Securities are as follows:

The robot training industry has evolved from an industrial customization phase primarily based on physical prototype testing to a virtual-physical integration and AI-driven generalized training paradigm.

In the early stages, robot training was mainly focused on industrial scene applications, relying heavily on physical equipment for specialized training in single scenarios, such as industrial robot operation training in the 1980s. With the advancement of science and technology and breakthroughs in AI and simulation technology, robot training has gradually shifted to virtual environments, achieving iterative optimization through technologies such as GPU acceleration and multimodal data synthesis. Currently, the robot industry is moving towards a full-scene coverage phase driven by embodied intelligent large models, with the training of robot brains and cerebellums becoming increasingly prominent as important tools for enhancing robot intelligence.

Generative AI is reshaping the training paradigm, significantly improving data generation efficiency.

Generative AI allows users to quickly generate new content based on various inputs. The inputs and outputs of these models can include text, images, sounds, animations, 3D models, or other types of data. With the support of generative models, the data required for AI training is gradually shifting from "primarily collected" to "primarily generated," and generative AI can utilize different learning methods (including unsupervised learning or semi-supervised learning) for training, making it easier and faster to build foundational models using large amounts of unlabeled data. For example, models like NVIDIA Dream Gen and Qunke Technology's Spatial LM can generate diverse training data from a small number of samples, reducing data acquisition costs by 80%. In industrial welding scenarios, only 100 real weld images are needed to generate 100,000 training samples. Furthermore, zero-shot and few-shot learning will also become more prevalent, and the combination of generative models with large language models will promote zero-shot generalization of robot tasks.

Physical AI is reshaping the underlying logic of robot training: shifting from "empiricism" reliant on real data to "rationalism" based on physical laws.

NVIDIA has built a complete ecosystem from cloud training to edge deployment through a full-stack technology solution, promoting physical AI from laboratories to industrial, medical, and household scenarios. In the future, as embodied intelligent large models combine with edge computing, robots will gradually penetrate various fields of human activity, ultimately achieving the ultimate goal of "thinking like humans and executing more efficiently than humans." In this revolution of the fusion of physics and intelligence, robots need a deeper understanding of the physical world to cope with complex real-world scenarios, becoming partners that help humans practically solve complex problems Risk Warning: Technical route iteration, data compliance risks, humanoid robot development not meeting expectations