Imagine stepping into a completely unfamiliar building. You don't have a map, but you intuitively know how to explore, find your way around obstacles, and eventually locate the exit. Giving robots this same ability to reason about space and plan their movements is a major challenge in AI. New research explores how deep learning can empower embodied agents—like robots—to navigate unknown environments. The key lies in teaching AI to build its own understanding of the world, much like we do. This is achieved through differentiable planning, a technique where the AI learns a model of how its actions affect the environment and what rewards (like reaching a goal) it can expect. Instead of relying on pre-programmed rules, the AI learns from experience, figuring out how to avoid obstacles and efficiently reach targets. One innovation is the Collision Avoidance Long-Term Value Iteration Network (CALVIN), which learns to plan longer sequences of movements, improving its ability to tackle complex spaces like mazes. CALVIN doesn’t just react to what it immediately sees; it develops an internal “map” of rewards and values, helping it anticipate future outcomes. This approach has shown impressive results in simulated 3D environments, allowing robots to navigate unfamiliar virtual spaces with remarkable success. Even more exciting, the research extends these principles to real-world robot navigation, leveraging images taken from a robot exploring indoor spaces. While challenges remain in making these systems completely robust and safe for practical deployment, this research signifies a leap toward robots that can seamlessly move and interact within our world, much like humans do.