The latest boom in robotics represents a revolution in the way machines have learned to interact with the world. Roboticists used to dream big but build small; now they’re betting on AI models that can learn from vast data sets and adapt to different environments.

Imagine a pair of robot arms folding clothes: it’s not just about writing rules, but letting the machine try countless techniques through trial and error. This new approach has seen companies and investors pour $6.1 billion into humanoid robots in 2025, four times what was invested in 2024.

The arrival of ChatGPT in 2022 catalyzed this boom: trained on vast amounts of text, these models can predict the next action a machine should take, issuing dozens of motor commands every second. This conceptual shift seems to work for any helpful robot, from talking assistants to social robots and even complex tasks.

However, early attempts like Jibo, an armless, faceless lamp-like robot, showed that better language capabilities are essential. Competing against established voice technologies like Siri and Alexa, Jibo’s early limitations were clear: scripted conversations couldn’t match the engaging and unpredictable nature of AI-generated ones. Now, while these new models can handle a wider range of tasks, they come with their own risks.

OpenAI's Dactyl is a robot hand trained virtually to manipulate objects; it’s a testament to how well simulations and real-world differences must be aligned for robots to truly adapt. domain randomization offers solutions, but the journey from simulation to reality remains fraught with challenges.