Artificial intelligence has undoubtedly transformed the landscape of human interaction with technology, showcasing capabilities that often feel eerily humanlike. From generating coherent text to creating audio and video content, contemporary AI systems have marked significant progress. Nevertheless, the real challenge lies not in generating digital content but in applying this intelligence effectively in the physical world. The reality is that despite impressive advancements, AI has largely remained confined to two-dimensional technology, with most algorithms struggling to adapt to the complexities of our three-dimensional reality.
Take, for instance, the quest for fully autonomous vehicles. The consistently elusive goal of creating a safe and reliable self-driving car encapsulates the larger issue—AI systems, while sophisticated, lack a fundamental understanding of real-world physics. They frequently make nonsensical errors, which can lead to catastrophic results. The question then arises: how do we bridge this critical gap? Emerging research proposes a solution embodying what can be termed “physical intelligence.”
Physical intelligence is not merely another iteration of AI; it represents a profound rethinking of how machines perceive and interact with their environments. Unlike traditional AI, which relies heavily on data interpretation, physical intelligence emphasizes an understanding of the physical principles that govern real-world dynamics. By leveraging awareness of cause-and-effect relationships, these systems become capable of understanding their surroundings in a more fluid and functional way.
The implications of this new paradigm are vast. In our pursuit of a more responsive technological ecosystem, the fusion of AI’s cognitive abilities with the mechanics of robotics can lead to machines that are not only intelligent but adaptable. My research team at MIT is at the forefront of this innovation, developing what we term “liquid networks.” These networks are designed to continually learn and adapt, mimicking human cognitive flexibility rather than being restricted by their initial training.
To illustrate the transformative potential of liquid networks, we conducted an experiment using two drones: one governed by conventional AI and the other by a liquid network. Tasked with the mission of locating objects within a forest, both drones excelled under the previously defined conditions. However, when exposed to unconventional scenarios, such as a winter landscape or an urban environment, only the drone operated by the liquid network achieved success. This demonstrated a remarkable distinction between static AI systems and innovative ones that can evolve through experience, a trait typically associated with human learning.
Moreover, physical intelligence systems transcend mere data interpretation and can execute elaborate commands derived from varied prompts. For instance, my lab has pioneered a system capable of designing and 3D-printing robots based on simple instructions like “create a walking robot” or “build a robot that can grip.” This capability hints at a future where user-interface commands can seamlessly translate into tangible actions within physical space, dissolving barriers between the digital and real worlds.
The advancements in this field are echoed across various laboratories and startups. For example, robotics company Covariant has devised chatbots resembling advanced models like ChatGPT that control robotic arms based on user instructions. Their considerable investment of over $222 million underscores the promise of intelligent robotic systems in practical applications such as warehouse sorting.
Notably, a team at Carnegie Mellon University recently showcased a robot capable of executing complex parkour maneuvers, demonstrating agility even with minimal sensory input. These developments indicate a pivotal shift where machines, through enhanced physical intelligence, can perform tasks previously considered too intricate or unpredictable for robotic execution.
As we look ahead, 2025 could herald a transformative era characterized by the emergence of devices across various domains—ranging from household amenities to large-scale infrastructure—that operate with an understanding of the commands imparted to them. Physical intelligence may soon redefine interaction dynamics, ushering in a phase of intuitive systems that learn to bridge the digital framework with our tactile reality.
The quest for a genuinely intelligent machine extends beyond programming and datasets; it involves creating systems capable of navigating the unpredictability of the real world with grace and precision. The future of AI lies not just in computation but in the acquisition of a nuanced, physical understanding of the world around us.