Recent advancements in AI have showcased models that generate text, audio, and video with surprising human-like capabilities. However, these algorithms have primarily remained confined to the digital domain, struggling to transition into our unpredictable physical world. This limitation is evident in the difficulties faced in developing reliable self-driving cars, where AI often lacks an understanding of physics and may produce unexpected errors.
As we move into 2025, a pivotal shift looms on the horizon: AI is set to break free from its digital shackles and begin functioning within our tangible environment. This transition involves reshaping how machines interpret their surroundings, merging AI’s digital intelligence with robotics’ mechanical capabilities. Termed "physical intelligence," this innovation will empower machines to navigate complex environments, adapt to changes, and make real-time decisions based on their understanding of the principles that govern the physical world.
For instance, at MIT, researchers are pioneering "liquid networks," models designed to exhibit physical intelligence. An illustrative experiment involved training two drones in locating objects within a forest—one equipped with traditional AI and the other with a liquid network. While both performed well under controlled conditions, only the liquid-network drone succeeded when faced with new challenges, such as finding objects in winter or urban settings. Unlike conventional AI systems, liquid networks demonstrate the ability to learn and evolve beyond their initial programming, akin to human adaptability.
Physical intelligence encompasses not only understanding commands but also executing them in real life. Researchers have created systems capable of designing and 3D-printing small robots in response to simple prompts, showcasing the ability of these intelligent machines to translate abstract instructions into physical actions.
Various labs are reporting breakthroughs in this field. Robotics startup Covariant is developing chatbots that can operate robotic arms based on user prompts, having secured substantial funding to deploy sorting robots in warehouses worldwide. Additionally, a team at Carnegie Mellon University has demonstrated a robot performing impressive parkour maneuvers using minimal visual input and reinforcement learning.
Thus, while recent years have been marked by the ascendance of text-based models, the upcoming era will be defined by physical intelligence. Expect to see a surge in devices—from robots to smart home technologies—capable of understanding and executing real-world tasks effectively.
