Embodied intelligence yet to see 'GPT moment'

Editor's note: Robotics company Unitree Robotics made its debut on the Shanghai Stock Exchange's Science and Technology Innovation Board on Wednesday, coinciding with the opening of the World Robot Conference in Beijing. ThePaper.cn, a Shanghai-based media outlet, spoke to Wu Zuxuan, vice-dean of the Institute of Trustworthy Embodied AI at Fudan University, on the development of China's embodied intelligence industry. Below are excerpts of the interview. The views don't necessarily represent those of China Daily.

A robot's journey begins with its physical form, which can be humanoid, quadruped, bipedal or shaped like a robotic arm. China holds a relatively strong position in this area. The country's core strengths are its robust industry and supply chains, which allow it to quickly source specific robot parts. Additionally, China's well-developed manufacturing sector enables the production of high-performance physical bodies of robots.

The strength of China's industry chain is also reflected in its ability to mass-produce and iterate rapidly. The country is able to develop the physical bodies in a short time, put them into mass production, and continuously upgrade the products based on the feedback from users.

Humanoid robots have been developed because most of today's infrastructure is designed for humans. It is unlikely that an entirely new infrastructure will be built for robots with completely different forms.

However, not all robots are required to be humanoid. For instance, robots in factories do not necessarily have to appear like humans. Large robotic arms are already accomplishing highly complex tasks in many automated factories.

Similarly, humanoid robots are not necessary for high-risk firefighting missions. In such scenarios, wheeled robots that move at a higher speed may be a better choice. In the future, robots will likely take diverse forms to perform different tasks in various scenarios.

Conventional robots mainly use grippers for operations. To perform more delicate tasks, robots need multifingered end effectors that mimic the coordinated movements of human hands. These dexterous hands should be durable, agile and capable of handling a wide range of tasks.

The development of large models for embodied intelligence is still in a nascent stage. We have yet to witness the "GPT moment" for embodied intelligence. The core challenge is scaling up data for embodied intelligence.

Large language models have access to massive volumes of internet data, which helps them perform diverse tasks. However, embodied intelligence suffers from insufficient data. Many questions remain about how embodied intelligence models interact with real-world environments, leaving considerable room for improvement.

Evaluating the performance of a large language model involves feeding it questions and assessing the responses it generates. In contrast, the test for evaluating embodied intelligence models is their ability to satisfactorily perform real-world tasks.

The core benchmark for the "GPT moment" of embodied intelligence will be when a single model can perform well across different kinds of robots, tasks and scenarios, demonstrating a strong ability to generalize.