Google’s new AI model lets robots learn in video game-style simulations

Google has revealed Genie 3, its newest AI “world model” that simulates realistic environments to help train robots and autonomous systems. Developed by Google DeepMind, Genie 3 takes a significant step toward the long-term vision of artificial general intelligence (AGI) by allowing AI agents to interact with complex, dynamic, and memory-rich environments.

Training robots in virtual warehouses

With Genie 3, Google envisions training autonomous machines like warehouse robots or delivery vehicles in digital environments that replicate real-world physics and behaviors. These simulated settings are not just static images—they evolve in real time, respond to text prompts, and support interactive events such as changing weather or introducing new objects.

“We expect this technology to play a critical role as we push toward AGI, and agents play a greater role in the world,” DeepMind said.

What makes Genie 3 different

Compared to its predecessor Genie 2, which allowed only 10–20 seconds of interaction, Genie 3 enables users and AI agents to explore simulated environments for several minutes. Even more impressively, it has short-term memory—if you turn away from a wall with a chalk drawing and turn back, the drawing is still there.

This consistency in visual memory is a game-changer. It creates a more believable and stable virtual space where AI agents can develop decision-making skills based on continuity—something crucial for future human-level AI applications.

Real-time world generation from simple prompts

Genie 3 can create environments instantly from text. For instance, typing “ski slope with deer” can produce an interactive snowy mountain scene populated by a herd of deer. The realism and responsiveness mimic video game experiences but are powered entirely by generative AI, not prebuilt assets.

Limitations and current availability

Despite the promise, Genie 3 is not yet ready for public use. Google is offering it as a limited research preview to a small group of academics and creators to evaluate potential risks. As of now, interactions are somewhat restricted, and readable text often only appears if it’s explicitly mentioned in the input description.

Still, Genie 3 represents a meaningful step forward in how AI can “learn” through simulation rather than relying solely on internet data.

“If you give a disembodied AI the ability to be embodied, albeit virtually, then the AI can explore the world, or a world – and grow in capabilities as a result,” said Andrew Rogoyski from the Institute for People-Centred AI.

The road to artificial general intelligence

World models like Genie 3 are seen by Google as essential tools on the path to AGI. These models let AI practice decision-making and task execution in a controlled but realistic space—learning from virtual mistakes before making real ones.

“To achieve flexible decision-making, robots need to anticipate the consequences of different actions to choose the best one to execute in the physical world,” said Prof Subramanian Ramamoorthy, a robotics expert at the University of Edinburgh.

While AGI is still a theoretical goal, advances like Genie 3 show that AI is inching closer to systems that can not only think like humans but also interact with the world as we do—step by simulated step.