Recent AI the models are surprisingly human in their ability to generate text, audio, and video when prompted. However, until now, these algorithms have remained largely relegated to the digital world rather than the three-dimensional physical world in which we live. In fact, whenever we try to apply these models to the real world, even the most sophisticated ones struggle to work adequately. – just think, for example, how difficult it has been to develop safe and reliable self-driving cars. Although they are artificially intelligent, not only do these models simply have no understanding of physics, but they also often hallucinate, leading them to make inexplicable mistakes.
But this is the year AI finally goes move from the digital world to the real world we live in. Expanding AI beyond its digital boundaries requires rethinking the way machines think, merging the digital intelligence of AI with the mechanical prowess of robotics. This is what I call “physical intelligence,” a new form of intelligent machine capable of understanding dynamic environments, coping with unpredictability, and making decisions in real time. Unlike the models used by standard AI, physical intelligence is rooted in physics; in understanding fundamental principles of the real world, such as cause and effect.
Such features allow physical intelligence models to interact and adapt to different environments. In my research group at MIT, we develop models of physical intelligence that we call liquid networks. In one experiment, for example, we trained two drones – one driven 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 when tasked to do exactly what they were trained to do, when asked to locate objects in different circumstances (in winter or in an urban environment) , only the liquid network drone succeeded in its task. This experiment showed us that, unlike traditional AI systems that stop evolving after their initial training phase, liquid networks continue to learn and adapt from experience, just as do humans.
Physical intelligence is also capable of physically interpreting and executing complex commands derived from text or images, bridging the gap between digital instructions and real-world execution. For example, in my lab we have developed a physically intelligent system that, in less than a minute, can iteratively design and then 3D print small robots based on prompts such as “robot that can move forward” or “ robot that can grasp objects.
Other laboratories are also making significant advances. For example, robotics startup Covariant, founded by UC-Berkeley researcher Pieter Abbeel, is developing chatbots, similar to ChatGTP, that can control robotic arms when prompted. They have already secured more than $222 million to develop and deploy sorting robots in warehouses around the world. A team from Carnegie Mellon University also recently demonstrated that a robot with a single camera and imprecise actuation can perform dynamic and complex parkour moves, including jumping over obstacles twice as high and through gaps twice as long, using a single network neural trained via reinforcement learning.
If 2023 was the year of text-image and 2024 of text-video, then 2025 will mark the era of physical intelligence, with a new generation of devices – not just robots, but everything from power grids to smart homes. – who can interpret what we tell them and perform tasks in the real world.
