The generation who grew up watching Star Wars may feel disappointed, as they still cannot see the shadow of C-3PO robots in their sight. When will humanoid robots with expected common sense and the ability to provide assistance to humans at home and in the workplace appear? The rapid development of artificial intelligence (AI) may soon fill this gap. The website of Nature recently reported that the "marriage" between AI and robotics technology may be a "two-way journey" that will push each other to new heights. AI instills common sense into robots, enabling them to handle various tasks like humans; Robots, on the other hand, utilize their physical experience to help AI leap to become a "universal AI". However, experts remind that this two-way journey still faces many challenges, including collecting massive and accurate data for robots to learn, dealing with unpredictable hardware, and solving security issues. Basic models help robots become smarter. Many AI scientists in the field of robotics hope that in the future, robots can autonomously adapt to a wider range of environments. For example, robotic arms that can help customers select products, intelligent humanoid robots that can accompany the elderly, and so on. However, controlling such robots is not an easy task. From OpenAI to Google's Deep Thinking company, they are committed to embedding multifunctional learning algorithms (foundational models) into robots to behave more like humans during conversations. The core of these ideas is to instill common sense in robots to enable them to perform various tasks. Gerard Andrews, a marketing expert in robotics at NVIDIA, emphasized that robotics technology is currently at a critical moment of change. The company launched a universal AI model specifically designed for humanoid robots in March. Google's Deep Thinking company has created one of the most advanced robot basic models, the RT-2, which can recognize photos of celebrities, although the celebrity has not appeared in previous training. Researchers optimistically believe that basic models can help scientists develop general-purpose robots that can assist human labor. In February of this year, humanoid robot company raised $675 million to apply OpenAI developed language and visual models in universal humanoid robots. A video shows a robot giving a person an apple in response to their request to eat. A variety of data makes robots "more experienced" Although chat robots are being trained with billions of words from the Internet, the data set of robot activities is far from enough. Data collection or a solution. Alexander Kahazetsky, a robotics research expert at Stanford University in the United States, and his colleagues have created the open-source dataset DROID2, which collects approximately 350 hours of video data from the Franka Panda 7DoF robotic arm manufactured by Franka Robotics in Germany, including records of hundreds of environments such as bathrooms, laundry rooms, and bedrooms. This diversity helps robots perform tasks better in new environments. California AI company Covariant is also working to expand the scale of robot data. The company began collecting data from 30 types of robotic arms around the world in 2018, and its Robot Basic Model 1 (RFM-1) not only collects video data but also sensor readings. This type of data is crucial for improving the robot's task execution ability, such as letting the robot know how to avoid hitting bananas. In addition, many experts are working on building 3D virtual reality environments with the aim of training robots through these environments. Both Meta and Nvidia have invested significant resources in the simulation field to expand the scale of robot data and establish complex simulation worlds. Through these platforms, robots can accumulate data equivalent to years of practical experience in just a few hours. Hardware and security issues need to be considered. Many robotics experts point out that although AI has given robots a smarter "brain", its further development still faces many challenges. Robots are very complex and their hardware is easily damaged. Harold Soh, a human-computer interaction expert at the National University of Singapore, said that robots may require a large amount of other types of data, such as tactile or proprioceptive (body position in space) data. This type of dataset currently does not exist, but it is necessary for humanoid robots to work efficiently. Security must be carefully considered. Large language models have been proven to generate errors, biases, and may also deceive humans or robots into doing some "bad things.". In view of this, R&D personnel need to implant some rules into the robot AI model. Robots undoubtedly still require a lot of manual supervision. Despite certain risks, the two-way journey between AI and robots is still worth looking forward to. By combining with robots and increasing interaction with the real world, AI is expected to surpass learning modes to make predictions, truly understand and reason about the world. (Lai Xin She)