Skip to main content

Scientists Discover Way to Send Information into Black Holes Without Using Energy

3 comments

New AI Breakthrough Allows Robots To See The Future. Here's How

Imagine a world where robots don’t just follow instructions—they think ahead, picturing the consequences of their actions before moving a single joint. That future is closer than you might think. A team led by Yilun Du, a researcher at the Kempner Institute, has unveiled a groundbreaking artificial intelligence (AI) system that allows robots to “envision” their next steps using video, a development that could transform how machines navigate and interact with the physical world.

This breakthrough, detailed in a preprint on arXiv and explained in the Kempner AI blog, represents a major shift in how researchers approach robot learning. Instead of relying solely on language-based instructions or trial-and-error learning, the system uses video to train robots on how the world behaves, giving them the ability to anticipate outcomes in ways that were previously impossible.

From Words to Vision: A New Era in Robotic Intelligence

For years, robotics researchers have relied on vision-language-action (VLA) systems, which are essentially foundation models for robots. These systems combine the ability to see, understand, and move, giving robots general-purpose skills. In theory, this reduces the need to retrain a robot every time it encounters a new environment or task.

However, even the most advanced VLA systems face challenges. Many depend heavily on large language models (LLMs) to translate human instructions into robot actions. While this approach works well for simple commands, it struggles to teach robots how to generalize knowledge when they face entirely new situations.

Du explains, “Language contains little direct information about how the physical world behaves. Our idea is to train a model on a large amount of internet video data, which contains rich physical and semantic information about tasks.”

Instead of reading instructions, the robot learns by watching the world—essentially learning physics and movement patterns through video.

Teaching Robots to Imagine the Future

At the heart of this innovation is a concept called a “world model”. Using the Kempner AI Cluster, one of the most powerful academic supercomputers, Du and his team encoded massive amounts of internet video data into an internal representation of the physical world. This allows the robot to generate short, imagined video clips of potential future scenarios before taking action.

“The video generation capabilities learned from internet data help transfer knowledge to the robot foundation model,” Du says. “The idea is to synthesize videos showing how a robot should act in a new task.”

In practical terms, this means the robot can simulate multiple possible outcomes and choose the most promising course of action. It can anticipate the result of picking up a cup, turning a door handle, or stacking objects—even in unfamiliar environments.

The research team demonstrated that this “visual imagination” lets robots perform a wide range of tasks in environments they haven’t seen before. Instead of trial and error, the robot can foresee the effects of its actions, much like a human would mentally rehearse a sequence of movements before doing them.

The Challenge of Physical Intelligence

Du emphasizes that this work highlights a broader truth about intelligence. Humans often associate intelligence with abstract problem-solving, like chess or math. But true physical intelligence—the ability to navigate and manipulate the world—requires far more than abstract reasoning.

“Physical intelligence is challenging because of the enormous diversity of environments,” Du says. “As you go through life—from your home to the outdoors, to a museum, underwater, or in the sky—you can still perceive, adapt, and navigate effectively, no matter how different the surroundings are.”

Another challenge is temporal dependency. Unlike a chess move, which is a single decision, physical tasks often require coordinating multiple actions over time. A robot must execute a sequence of steps in the correct order to achieve success, such as turning a valve, moving an object, and adjusting its grip—all while responding to changes in its environment.

By using video, Du’s team is giving robots the ability to anticipate not just one move, but a chain of consequences. The model predicts how the world will evolve visually, aligning more closely with the physics of the real world than language-based instructions ever could.

Toward Robots That Understand Like Living Creatures

One of the most fascinating aspects of this research is how it mirrors biological intelligence. Humans didn’t evolve to follow written instructions; we evolved to interact physically with our surroundings. Our intelligence is built on millions of years of motor learning, perception, and adaptation.

Du notes, “It seems that the right way to develop intelligent robots is not by training models primarily on language information. Language didn’t teach us how to interact in the physical world.”

By grounding AI in sensory patterns and physical interactions, the researchers are nudging robotics toward a more natural, life-like understanding of the world. In essence, these robots are beginning to see, imagine, and act in ways that resemble how animals, including humans, interact with their environments.

Looking Ahead: Long-Term Planning and Memory

The next step for Du and his collaborators is linking visual imagination with long-term planning and memory. For instance, if a robot is asked to navigate a house, it must not only plan a route but also remember past experiences: which rooms are cluttered, which doors open automatically, and where objects are typically located.

“Right now, many of our tasks are in relatively static environments, where the robot can pick up objects and interact without much change,” Du says. “But in more dynamic settings, a robot needs to account for things like object weight or changing conditions. Exploring how to handle those kinds of physical dynamics is another exciting challenge.”

By extending the system to handle long-term goals and dynamic environments, robots could one day navigate real-world spaces as effortlessly as humans do. This would open up possibilities in areas ranging from household assistance and disaster response to advanced manufacturing and exploration in hazardous terrains.

Implications for AI and Robotics

This research represents a paradigm shift in robotics. Instead of thinking about robots as purely computational machines executing coded instructions, we can imagine them as entities capable of predictive perception—machines that can mentally simulate the world before acting.

The use of video-based learning also highlights the importance of rich sensory data for AI. By observing how the world behaves, robots can learn physical rules, causal relationships, and task structures that are difficult to encode in language or symbolic instructions.

In other words, teaching robots to imagine may be the key to creating AI that is more generalizable, adaptable, and human-like.

Conclusion

Yilun Du and his team at the Kempner Institute have taken a major step toward robots that understand the world not through words, but through vision and imagination. By leveraging internet videos, supercomputing power, and advanced AI models, they’ve created machines that can anticipate future scenarios, simulate possible actions, and adapt to unfamiliar environments.

This approach challenges the traditional view of intelligence, emphasizing physical adaptability and temporal planning over abstract reasoning alone. It moves robotics closer to a biological model of understanding, one that is grounded in perception, action, and anticipation.

The next frontier is even more ambitious: robots that can remember, plan long-term, and adapt dynamically in a changing world. If successful, these systems could revolutionize industries, transform human-robot interaction, and bring us closer to machines that truly think ahead, like living creatures navigating their environment.

The research, titled “Large Video Planner Enables Generalizable Robot Control”, is a glimpse into a future where robots do more than follow orders—they imagine, predict, and understand the physical world in ways that are both flexible and astonishingly human-like.

Reference: Boyuan Chen et al, Large Video Planner Enables Generalizable Robot Control, arXiv (2025). DOI: 10.48550/arxiv.2512.15840

Comments

Post a Comment

Popular

Scientists Discover Way to Send Information into Black Holes Without Using Energy

For years, scientists believed that adding even one qubit (a unit of quantum information) to a black hole needed energy. This was based on the idea that a black hole’s entropy must increase with more information, which means it must gain energy. But a new study by Jonah Kudler-Flam and Geoff Penington changes that thinking. They found that quantum information can be teleported into a black hole without adding energy or increasing entropy . This works through a process called black hole decoherence , where “soft” radiation — very low-energy signals — carry information into the black hole. In their method, the qubit enters the black hole while a new pair of entangled particles (like Hawking radiation) is created. This keeps the total information balanced, so there's no violation of the laws of physics. The energy cost only shows up when information is erased from the outside — these are called zerobits . According to Landauer’s principle, erasing information always needs energy. But ...
3 comments
Read more

Black Holes That Never Dies

Black holes are powerful objects in space with gravity so strong that nothing can escape them. In the 1970s, Stephen Hawking showed that black holes can slowly lose energy by giving off tiny particles. This process is called Hawking radiation . Over time, the black hole gets smaller and hotter, and in the end, it disappears completely. But new research by Menezes and his team shows something different. Using a theory called Loop Quantum Gravity (LQG) , they studied black holes with quantum corrections. In their model, the black hole does not vanish completely. Instead, it stops shrinking when it reaches a very small size. This leftover is called a black hole remnant . They also studied something called grey-body factors , which affect how much energy escapes from a black hole. Their findings show that the black hole cools down and stops losing mass once it reaches a minimum mass . This new model removes the idea of a “singularity” at the center of the black hole and gives us a better ...
Post a Comment
Read more

How Planetary Movements Might Explain Sunspot Cycles and Solar Phenomena

Sunspots, dark patches on the Sun's surface, follow a cycle of increasing and decreasing activity every 11 years. For years, scientists have relied on the dynamo model to explain this cycle. According to this model, the Sun's magnetic field is generated by the movement of plasma and the Sun's rotation. However, this model does not fully explain why the sunspot cycle is sometimes unpredictable. Lauri Jetsu, a researcher, has proposed a new approach. Jetsu’s analysis, using a method called the Discrete Chi-square Method (DCM), suggests that planetary movements, especially those of Earth, Jupiter, and Mercury, play a key role in driving the sunspot cycle. His theory focuses on Flux Transfer Events (FTEs), where the magnetic fields of these planets interact with the Sun’s magnetic field. These interactions could create the sunspots and explain other solar phenomena like the Sun’s magnetic polarity reversing every 11 years. The Sun, our closest star, has been a subject of scient...
Post a Comment
Read more