Artificial intelligence (AI) has the capability to find recipes or generate images, but it struggles with physical tasks like hanging artwork or cooking. Chelsea Finn, an engineer and researcher at Stanford University, aims to change this, believing AI could usher in a new era in robotics. Finn, who co-founded a company that has showcased a versatile AI robot capable of folding laundry, emphasizes the need for software enabling robots to function intelligently across various scenarios. While advancements such as Google’s recent AI-powered lunch-packing robot illustrate potential, the research community is divided on whether generative AI can revolutionize robotics as it has online tasks. Critics, like UC Berkeley's Ken Goldberg, caution that robots won’t achieve their science fiction potential overnight due to the complexities involved. Historically, the perception of robotics has often outpaced reality; robots excel in controlled, repetitive tasks but struggle with unpredictable environments. Finn's lab is exploring solutions through a project called "OpenVLA, " which aims to create a system combining vision, language, and action, resembling the working of the human brain. This innovative approach allows for training robots to perform tasks through demonstration instead of detailed programming. Stanford's Moo Jin Kim illustrates this process, where a human “puppeteers” the robot using joysticks to teach it tasks through repetitive practice.

As the robot learns from these demonstrations, it begins to replicate the actions independently. Physical Intelligence, the company co-founded by Finn, seeks to expand this training model, envisioning robots capable of adapting to various simple tasks such as making sandwiches. Finn argues that developing generalist systems could prove more effective than specialized ones, though gathering the necessary training data presents challenges. Skeptics like Ken Goldberg believe that building a comprehensive dataset for physical tasks will take significantly longer than for chatbots due to the complexities of real-world data. Similarly, MIT’s Pulkit Agrawal favors simulations as a way to accelerate data collection by allowing robots to practice repeatedly in virtual environments, although such simulations often fail to capture the unpredictable nature of the real world. Deeper issues also persist, as highlighted by Carnegie Mellon’s Matthew Johnson-Roberson, who believes that robots must tackle more complex spatial and temporal challenges than chatbots, making the data insufficient for refined task execution. Despite skepticism, many acknowledge that AI will transform robotics. Goldberg’s company, Ambi Robotics, has developed an AI-driven system that enhances package sorting, significantly improving efficiency while still requiring traditional programming for execution. While Finn agrees that advancements are needed, she remains optimistic about AI-powered robots assisting in sectors facing labor shortages due to aging populations, although she doesn't foresee these robots replacing human labor for complex tasks completely.