AI Takes the controls: First “self-driving” telescope successfully observes the night sky

Artificial intelligence has already transformed fields ranging from healthcare and finance to logistics and autonomous vehicles. Now, AI is turning its attention to one of humanity’s oldest scientific pursuits: observing the stars.

Researchers from Northwestern University, the University of Chicago, and the U.S. Department of Energy’s Fermilab have successfully demonstrated what they describe as the first AI-driven telescope scheduling system, marking an important step toward autonomous astronomical observatories. The technology has already been tested on one of the world’s most productive astronomical facilities, showing that AI can make complex observing decisions in real time while adapting to changing conditions throughout the night.

The achievement could help astronomers make better use of scarce telescope time, accelerate scientific discovery, and offer valuable lessons for Canada’s rapidly growing AI and astronomy sectors.

The problem with traditional telescope scheduling

Observing the universe is far more complicated than simply pointing a telescope at an interesting celestial object. Every night, astronomers must weigh multiple factors before deciding where a telescope should focus. Cloud cover, atmospheric stability, moonlight brightness, object visibility, scientific priorities, and telescope availability all play a role in determining the most productive use of observing time.

The challenge is particularly significant because access to large telescopes is highly competitive. Researchers often wait months or even years for approved observation slots. If conditions deteriorate or a telescope is directed inefficiently, valuable scientific opportunities can be lost.

As Alex Drlica-Wagner, a professor of astronomy and astrophysics at the University of Chicago and scientist at Fermilab, notes, major telescopes are international scientific resources used by researchers around the world. Every minute of observing time carries significant value.

Historically, these scheduling decisions have depended heavily on the expertise of experienced astronomers. The new AI system aims to automate much of that process.

Training AI to think like an astronomer

Rather than programming the system with hundreds of specific rules developed over decades of telescope operations, the research team chose a different approach. They allowed the AI to learn by example.

The researchers trained a deep-learning model using 13 years of historical observations gathered through the Dark Energy Survey, a major astronomical project that used the Dark Energy Camera (DECam) mounted on the Víctor M. Blanco 4-meter Telescope in Chile.

The system was shown where the telescope was observing at a given moment and then asked to predict its next target. By repeatedly comparing predictions against decisions made by human astronomers, the AI learned the observational strategies experts use when making scheduling choices.

According to the researchers, the system was never explicitly taught rules relating to moonlight conditions, atmospheric quality, or image optimization. Instead, it learned these relationships independently from historical data.

This approach represents a growing trend in artificial intelligence: allowing machine learning systems to discover complex patterns that would be difficult to encode manually.

The real test came when the AI was deployed on an operational observatory. Using the Dark Energy Camera, a highly sophisticated 570-megapixel instrument, the researchers conducted two successful observing runs during the spring and summer of 2026 at the NSF Víctor M. Blanco Telescope at Cerro Tololo Inter-American Observatory in Chile.

The AI system generated observing plans and adapted those plans as conditions evolved throughout the night. Changes in weather or sky conditions that might otherwise require human intervention were incorporated automatically into revised observation schedules. For the initial deployment, the objective was modest but important: match human performance.

According to the research team, the AI scheduler achieved results comparable to those of experienced human operators. Having demonstrated this capability, the next stage of development is considerably more ambitious.

Researchers now hope to create systems capable of outperforming human schedulers by identifying observing strategies that people may never have considered.

Modern astronomy is entering an era of unprecedented data generation. Next-generation facilities such as the NSF-DOE Vera C. Rubin Observatory are expected to produce enormous volumes of astronomical data. Managing observations and coordinating follow-up investigations will become increasingly challenging. AI may be uniquely suited to this environment.

Intelligent scheduling systems can rapidly process changing conditions, assess competing priorities, and optimize telescope usage in ways that would be difficult for human operators to perform consistently over long periods. The result could be more efficient scientific operations and increased research productivity.

Importantly, automation can also free astronomers from routine operational decisions, allowing them to focus more attention on scientific interpretation and discovery.

As Drlica-Wagner suggests, removing some of the technical burden of observation planning may enable researchers to spend more time addressing fundamental scientific questions.

A Canadian perspective

While the project was conducted in the U.S. and Chile, the implications are highly relevant for Canada. Canada has emerged as one of the world’s leading centres for artificial intelligence research, thanks in large part to organizations such as Mila in Montréal, the Vector Institute in Toronto, and Amii in Edmonton. Canadian researchers have played a major role in advancing machine learning techniques now used worldwide.

Canada also has a distinguished tradition in astronomy and astrophysics. Canadian scientists contribute to international telescope projects, cosmology research, exoplanet studies, and observational astronomy initiatives around the globe.

The convergence of AI and astronomy represents a particularly exciting opportunity. Projects such as the AI Institute for the Sky (SkAI), which supported the telescope scheduling work, illustrate how machine learning can become an active partner in scientific discovery. Similar approaches could eventually be applied to observatories used by Canadian researchers or incorporated into future international astronomy collaborations involving Canadian institutions.

There are also broader industrial implications. The techniques developed for autonomous telescope operations have similarities to challenges faced in autonomous vehicles, smart manufacturing systems, robotics, and remote sensing platforms. Advances in one sector often generate innovations that benefit many others.

The successful deployment of the AI scheduling system points toward a future where observatories become increasingly autonomous. Rather than relying on continuous human oversight, future telescopes may monitor environmental conditions, select optimal targets, coordinate with other observatories, and adjust observing strategies automatically.Such systems could become especially important in remote environments where staffing is difficult or expensive. They may also prove essential as the number of astronomical surveys continues to grow and observational complexity increases.