To Interact With the Real World, AI Will Gain Physical Intelligence

latest AI The model’s ability to generate text, audio, and video when prompted is strikingly similar to humans. However, until now, these algorithms have remained largely limited to the digital world rather than the physical three-dimensional world we live in. It’s also difficult to fully execute—think, for example, how challenging it is to develop safe and reliable self-driving cars. Although artificial intelligence, these models not only do not understand physics at all, but also often produce hallucinations, which causes them to make inexplicable mistakes.

However, this year, artificial intelligence will finally From the digital world to the real world we live in. Extending AI beyond the digital boundary requires re-engineering the way machines think, fusing the digital intelligence of AI with the mechanical capabilities of robots. This is what I call “physical intelligence,” a new class of intelligent machines capable of understanding dynamic environments, coping with unpredictability, and making decisions on the fly. Unlike the models used by standard artificial intelligence, physical intelligence is rooted in physics. Understand fundamental principles of the real world, such as cause and effect.

These capabilities allow physical intelligence models to interact and adapt to different environments. In a research group at MIT, we are developing models of physical intelligence that we call liquid networks. For example, in one experiment, we trained two drones (one operated by a standard AI model and the other by a liquid network) to locate objects in a forest during the summer using data captured by human pilots. While both drones performed equally well on their training missions, when they were asked to locate objects in different conditions (winter or urban environments), only the LiquidNet drone successfully completed the task. This experiment shows us that, unlike traditional artificial intelligence systems that stop evolving after an initial training phase, liquid networks continue to learn and adapt from experience, just like humans.

Physical intelligence is also able to interpret and actually execute complex commands derived from words or images, bridging the gap between digital instructions and real-world execution. For example, in my lab, we developed a physical intelligence system that can iterate designs based on prompts such as “a robot that can walk forward” or “a robot that can grasp” and then 3D Print small robots.

Other labs have also made major breakthroughs. For example, Covariant, a robotics startup founded by UC Berkeley researcher Pieter Abbeel, is developing a chatbot (similar to ChatGTP) that can control a robotic arm when prompted. They have raised more than $222 million to develop and deploy sorting robots in warehouses around the world. A team from Carnegie Mellon University recently also proved Equipped with just a camera and imprecise actuation, a robot can perform dynamic and complex parkour movements, including jumping up obstacles twice as high and spanning gaps twice as long, using a single neural network trained through reinforcement learning.

If 2023 was the year of words to images, and 2024 the year of words to film, then 2025 will mark the age of physical intelligence, with a new generation of devices that are not just robots but everything from power grids to smart homes. It can interpret what we tell them and perform tasks in the real world.