How robots learn: A brief, contemporary history

In 2025, a quiet revolution unfolded in garages, research labs, and corporate innovation centers around the world. Robots, once confined to repetitive tasks on factory floors or vacuuming our living rooms, began to learn—truly learn—how to interact with the physical world. The dream of humanoid assistants once reserved for science fiction is now edging toward reality, not because engineers finally perfected mechanical precision, but because machines are now capable of learning from experience, much like humans do. This shift marks a fundamental transformation in robotics: from rigid programming to adaptive intelligence.

For decades, roboticists chased the dream of building machines that could match the dexterity, adaptability, and intuition of the human body. But despite grand visions of C-3PO-like companions, the reality was often a robotic arm bolted to an assembly line, repeating the same motion thousands of times with unerring accuracy. These machines were marvels of engineering, but they were not intelligent. They couldn’t fold a shirt, pick up a coffee cup without spilling, or navigate a cluttered kitchen. Their intelligence was brittle—hard-coded by engineers who had to anticipate every possible scenario. If a screw was slightly misaligned or a fabric wrinkled unexpectedly, the robot would freeze, confused by the deviation from its script.

Today, that paradigm is crumbling. The explosion of investment—$6.1 billion poured into humanoid robots in 2025 alone, quadruple the previous year’s funding—signals a new era. The money isn’t just chasing better motors or stronger materials; it’s betting on a new kind of learning. Robots are no longer being taught; they’re being trained. And the methods mirror those used to develop artificial intelligence in other domains, from language to vision. The key breakthrough? A shift from rule-based programming to data-driven learning, powered by vast neural networks and real-world experience.

The Old Way: Programming Every Possibility

In the early days of robotics, engineers approached machine intelligence like architects designing a building: every beam, joint, and load had to be calculated in advance. To teach a robot to fold clothes, for example, programmers had to write thousands of lines of code specifying how to detect fabric type, identify seams, calculate folding angles, and adjust for wrinkles. If the shirt was inside out, the robot needed a separate subroutine. If the fabric was silk instead of cotton, another set of rules applied. The complexity grew exponentially with each new variable.

This method, known as explicit programming, required human experts to anticipate every possible scenario. It was like trying to write a dictionary that includes every word, every slang term, and every regional dialect before someone even speaks. The robot could only succeed within the narrow boundaries defined by its code. Any deviation—a crumpled sleeve, a missing button, a dog sleeping on the laundry pile—would cause it to fail. This approach worked well in controlled environments like factories, where conditions are predictable and repeatable. But in the messy, unpredictable world of homes and public spaces, it was hopelessly inadequate.

Even the most advanced robots of the 2010s, like those developed by Boston Dynamics, relied heavily on pre-programmed behaviors. Their iconic backflips and parkour moves were not learned through trial and error but choreographed by engineers using physics simulations and motion capture. While visually stunning, these robots lacked true adaptability. They couldn’t learn to open a new type of door or carry a fragile object without explicit instructions.

The Simulation Breakthrough: Learning by Doing

Around 2015, a new approach began to emerge—one inspired by how children learn to walk or play. Instead of programming every action, researchers started building digital twins: highly detailed simulations of robots and their environments. In these virtual worlds, robots could practice tasks millions of times, trying different strategies and receiving feedback in the form of rewards or penalties.

This method, known as reinforcement learning, mirrors the way AI mastered games like chess and Go. A robot arm in simulation might attempt to fold a virtual shirt 100,000 times. Most attempts fail—the fabric tears, the folds are uneven, the robot drops the shirt. But each failure provides data. The system adjusts its neural network, slowly improving its technique. Over time, it discovers efficient, robust strategies that no human programmer could have devised.

One striking example comes from researchers at UC Berkeley, who trained a robot to fold towels using only simulation. After 5,000 virtual attempts, the robot achieved a 90% success rate—without ever touching real fabric. When transferred to a physical robot, it performed nearly as well, thanks to techniques that bridge the “sim-to-real” gap. This marked a turning point: robots were no longer just executing commands; they were learning from experience.

The Language Model Revolution: From Words to Actions

The real game-changer arrived in 2022 with the release of ChatGPT. This large language model (LLM) didn’t just generate text—it demonstrated an uncanny ability to understand context, reason, and adapt. Suddenly, the idea of training models on vast amounts of data to perform complex tasks seemed not just possible, but inevitable.

Roboticists quickly realized they could apply the same principle. Instead of training on text, they fed multimodal data—images, sensor readings, joint positions, and motor commands—into neural networks. These models learned to predict the next best action, just as LLMs predict the next word in a sentence. A robot watching a human fold clothes could analyze the sequence of movements and replicate them. It didn’t need to understand fabric physics; it just needed to mimic the pattern.

This approach, known as behavioral cloning, has enabled robots to perform tasks they’ve never been explicitly programmed to do. For instance, a robot trained on thousands of hours of human kitchen activity can now pour coffee, load a dishwasher, or chop vegetables—not because it was told how, but because it learned from observation. The model acts like a digital apprentice, absorbing patterns and refining its technique over time.

Learning in the Wild: Robots That Improve by Doing

Perhaps the most radical shift in modern robotics is the move toward deployment-based learning. Instead of waiting until a robot is perfect, companies are releasing early versions into real-world environments—homes, hospitals, warehouses—where they learn by interacting with people and objects.

This “learn by doing” philosophy mirrors how startups use minimum viable products (MVPs) to test ideas. A robot might start with limited capabilities—say, folding only T-shirts—but as it encounters new fabrics, shapes, and challenges, it updates its model. Data from thousands of robots is aggregated, analyzed, and used to improve the entire fleet. It’s a form of collective intelligence, where every robot contributes to a shared knowledge base.

Tesla’s Optimus robot, for example, is being tested in factories where it performs simple tasks like sorting parts. Each interaction feeds back into its neural network, allowing it to adapt to variations in lighting, object placement, and human behavior. Similarly, Amazon is deploying robot assistants in select fulfillment centers, where they learn to navigate dynamic environments and handle unexpected obstacles.

This approach accelerates learning far beyond what’s possible in simulation. Real-world data is messy, unpredictable, and rich with edge cases—exactly what robots need to become truly adaptable.

The Human Touch: Why Robots Still Need Us

Despite these advances, robots are not yet autonomous in the way humans are. They still require human oversight, data labeling, and occasional intervention. But the relationship is evolving. Instead of programming robots line by line, engineers now act more like coaches—curating data, designing reward systems, and guiding learning processes.

Moreover, robots are beginning to learn not just from data, but from human feedback. A user might correct a robot’s folding technique by demonstrating the right way, and the robot incorporates that correction into its model. This creates a feedback loop where humans and machines learn together.

Interestingly, this mirrors how children learn language and motor skills—not through isolated study, but through interaction, imitation, and correction. The future of robotics may not be about replacing humans, but about creating symbiotic partnerships where each enhances the other’s capabilities.

The Road Ahead: Challenges and Possibilities

The path to truly intelligent robots is still fraught with challenges. Energy efficiency, safety, and ethical concerns remain significant hurdles. A robot that learns too aggressively might break objects or injure people. Data privacy is another issue—robots collecting data in homes and hospitals must be carefully regulated.

Yet the momentum is undeniable. With billions in funding, rapid advances in AI, and a growing appetite for automation, the age of learning robots is just beginning. The dream of a robot that can fold your laundry, make your coffee, and chat with your kids is no longer fantasy. It’s a matter of time, training, and trust.

As one roboticist put it, “We’re not building machines that follow rules. We’re building machines that learn to care.” And in that shift—from code to compassion, from instruction to intuition—lies the future of robotics.

This article was curated from How robots learn: A brief, contemporary history via MIT Technology Review