The quest to determine what machines truly require to thrive in our physical environment has grown increasingly urgent. As modern technological frameworks empower machines to process information with unprecedented speed and sophistication, experts are forced to evaluate the boundaries of artificial cognition.

Intuition Versus Logical Cognition in Robotics

During the Minds in Motion Physical AI Summit, MIT associate professor and Eka co-founder Pulkit Agrawal outlined the fundamental modalities governing intelligence. He categorized human cognition into two distinct realms: System 1, which maps to instinctive intuition, and System 2, which drives logical reasoning. According to Agrawal, contemporary machines continue to lag behind biological capabilities largely because they remain confined to System 2 processing, creating performance bottlenecks.

Agrawal also highlighted the concept of the jagged frontier—a term popularized by Ethan Mollick—describing how machines excel remarkably in certain domains while struggling profoundly in others. To illustrate this phenomenon, he emphasized that not all data holds equal weight.

Not all data is born equal.

Demonstrating these principles, Agrawal showcased a robotic arm effortlessly tossing diverse small objects into a dish, executing real-time course corrections upon failure, and even delicately handling a fragile raspberry.

“The robot misses—it recovers from its mistake,” Agrawal explained, highlighting scenarios where a machine's internal System 1 proves entirely sufficient. “This robot is picking up, you know, 1500 objects per hour. And human speeds at this task are around 600 picks per hour… this is what I mean by system one: something which will be fast, something which will be reliable.”

The Pursuit of Superhuman Performance

Challenging engineers to raise their standards, Agrawal advocated for benchmarks that transcend human baselines. “(The goal should be) to be superhuman,” he stated regarding robotic development. “If robots can be more reliable, if we use them for surgeries, we could save more lives; if they go faster, we can do more manufacturing. The best products in human history have been superhuman. So what we are building is superhuman intelligence for the physical world, and that’s what Eka is trying to achieve.”

Drawing parallels to historical industrial milestones, Agrawal pointed to automobiles and calculators as prime examples of superhuman utility. Cars extended human travel distances and velocity beyond biological limits, while digital calculators revolutionized computational speed. Today, autonomous coding agents manage vast arrays of tasks far beyond individual human capacity. He asserted that the robotics industry will follow this exact trajectory.

Mastery Through Reinforcement Learning

To shatter existing performance ceilings, Agrawal emphasized the necessity of mastery achieved via reinforcement learning. By allowing algorithms to undergo trial and error across billions of simulated scenarios, these systems discover superior strategic execution that surpasses human capability—a phenomenon previously observed in elite chess and Go playing algorithms, as well as modern software development agents.

Scaling Through Simulation and Sandboxing

Another cornerstone of Agrawal’s presentation focused on the power of sandboxed training environments. Rather than relying solely on messy real-world data collection, systems are cultivated entirely within virtual simulations.

“Everything has been trained from simulation,” he noted. “And as we scale this up, it means that simulation scales at compute, which means there is no limit to how far we can scale it. And that is what we are super-excited about. And this is how we believe we can build the world’s best system one, to make robots become reliable, to make them be performant without the drudgery of real world data collection.”

As advancements accelerate across humanoid and quadrupedal designs from innovators like Tesla and Unitree, the inflection point for intelligent physical automation is rapidly approaching. The technological explosion is imminent, and the integration of smart machines into everyday environments is closer than many realize.