Researchers at Johns Hopkins University have developed GenEx, an AI system capable of imagining entire surroundings from a single image, mimicking human reasoning.
GenEx eliminates the need for physical exploration by creating a virtual world from one image, unlike older systems requiring costly, time-consuming environmental mapping.
“Humans imagine unseen surroundings using cues and knowledge,” says Professor Alan Yuille. GenEx replicates this reasoning for AI, making educated environmental predictions.
GenEx assigns probabilities to potential unseen environments instead of guessing a single possibility, enabling versatile and realistic mapping from limited visual data.
Applications of GenEx include disaster response, where rescue teams can explore hazardous sites remotely using just one surveillance image, avoiding human risk.
For example, imagining from another perspective, like a taxi driver stopping unexpectedly, helps GenEx predict unseen factors, useful for tasks like autonomous driving.
GenEx outperforms standard video generation benchmarks, enabling more accurate decision-making for both AI agents and human users in complex planning tasks.
The team integrated cognitive science, computer vision, and natural language processing into GenEx, advancing embodied AI toward humanlike intelligence.
Future plans for GenEx include incorporating real-world sensor data and dynamic scenes, improving immersive, realistic planning for various applications.
Explore GenEx through an interactive demo, with research details available on arXiv.org. This breakthrough marks a major step toward AI achieving humanlike reasoning.
Researchers at Johns Hopkins University have developed GenEx, an AI system capable of imagining entire surroundings from a single image, mimicking human reasoning.
GenEx eliminates the need for physical exploration by creating a virtual world from one image, unlike older systems requiring costly, time-consuming environmental mapping.
“Humans imagine unseen surroundings using cues and knowledge,” says Professor Alan Yuille. GenEx replicates this reasoning for AI, making educated environmental predictions.
GenEx assigns probabilities to potential unseen environments instead of guessing a single possibility, enabling versatile and realistic mapping from limited visual data.
Applications of GenEx include disaster response, where rescue teams can explore hazardous sites remotely using just one surveillance image, avoiding human risk.
GenEx improves navigation apps, trains autonomous robots, and enhances immersive gaming and VR by enabling reasoning-based exploration of synthetic virtual worlds.
Using a view, direction, and distance, GenEx agents can flexibly navigate and plan in their surroundings without needing to see everything physically.
GenEx ensures consistent environments through “spherical consistency learning,” guaranteeing seamless panoramic world modeling across exploration paths.
“We tested GenEx by having it navigate a closed path,” says Jieneng Chen. The model returned identical start and end views, proving its consistency.
GenEx empowers AI agents to make logical decisions with its "imagination-augmented policy," mimicking human reasoning in dynamic scenarios.
For example, imagining from another perspective, like a taxi driver stopping unexpectedly, helps GenEx predict unseen factors, useful for tasks like autonomous driving.
GenEx outperforms standard video generation benchmarks, enabling more accurate decision-making for both AI agents and human users in complex planning tasks.
The team integrated cognitive science, computer vision, and natural language processing into GenEx, advancing embodied AI toward humanlike intelligence.
Future plans for GenEx include incorporating real-world sensor data and dynamic scenes, improving immersive, realistic planning for various applications.
Explore GenEx through an interactive demo, with research details available on arXiv.org. This breakthrough marks a major step toward AI achieving humanlike reasoning.