The Dawn of Physical Intelligence: Bridging AI and the Tangible World

A new era of artificial intelligence, known as physical intelligence, is poised to revolutionize the interaction between AI and the real world. Expected to commence in 2025, this groundbreaking development promises an evolution from digital intelligence to a more dynamic integration of AI with mechanical capabilities. Rooted in the fundamental principles of physics, physical intelligence aims to transcend the limitations of traditional AI systems by continuously learning and adapting, much like humans.

Physical intelligence represents a fusion of digital smartness with the mechanical prowess inherent in robotics. Unlike traditional AI systems that cease evolving post-training, this new breed of intelligence thrives on perpetual adaptation through experience. A striking example of this capability was demonstrated by a team at Carnegie Mellon University. They showcased a robot equipped with one camera and imprecise actuation executing complex parkour movements. This feat was achieved using a single neural network trained via reinforcement learning, enabling the robot to leap onto obstacles twice its height and navigate gaps twice its length.

At the heart of physical intelligence lies the concept of liquid networks. These networks are designed to perform tasks under varying conditions, offering a stark contrast to their conventional AI counterparts. Researchers at MIT have been at the forefront of developing these liquid networks, which learn and adapt over time. In practical applications, robotics startup Covariant is pioneering chatbots capable of controlling robotic arms upon command. With over $222 million secured, Covariant aims to deploy sorting robots in warehouses worldwide, showcasing the potential of physically intelligent systems to execute real-world tasks effectively.

Moreover, physically intelligent systems boast remarkable capabilities, such as interpreting commands to design and 3D-print small robots in under a minute. This adaptability extends to drones as well. Researchers have successfully trained drones using both liquid networks and standard AI models to locate objects in dense forest environments during summer. Notably, drones equipped with liquid networks outperformed their peers by adeptly navigating diverse circumstances.