Tech • AI • Robotics
AI models can now help run physical science experiments
6/10TL;DR
A new Model Hardware Standard aims to let AI systems operate scientific instruments across labs, potentially cutting setup time from years to months and speeding work in neuroscience, microscopy and drug discovery.
KEY POINTS
A bottleneck in modern science
Building and debugging experiments often consumes most of a scientist’s effort, with estimates around 80% of time spent on hardware integration, software setup and calibration rather than on the core scientific question. The proposed solution targets that gap by giving AI a general way to communicate with physical devices and help run experiments.
Origins in a neuroscience lab
Neuroscientist Arco Bast spent about a year assembling a complex custom microscope to study how memories form in the brain in real time. The system relied on precise laser alignment and many separate components, each using different control methods. A command such as setting a beam to 50% power demonstrated that AI-mediated control could make disparate devices work together.
A general interface for instruments
Bast and collaborators developed what they call Model Hardware Standard, or MHS, as a device-agnostic layer between AI models and lab equipment. The goal is to let one system control microscopes, robotic arms and other tools without rewriting bespoke integrations for every machine. In effect, MHS is intended to translate among the different “languages” used by scientific hardware.
Safety controls are built in
Early tests on a robotic arm focused on defining safe operating limits before attempting more complex tasks. When the AI was asked to move beyond the permitted range, the system rejected the action rather than executing it. That safeguard is important in labs where a mistaken motion can damage expensive equipment or destroy delicate samples.
Rapid prototyping on unfamiliar devices
In one trial, the AI was given control of a robotic setup it had not previously operated and produced a working grasping behavior within minutes. Researchers described that speed as unusually fast compared with traditional lab automation, where building a custom workflow often requires substantial engineering work. The result suggested that a general AI-hardware interface could shorten development cycles dramatically.
Microscopy tests with Danaher and Leica
The project expanded to commercial lab equipment through work with Danaher and a Leica microscope. The AI had no specialized prior training for that exact application and initially tried different settings to obtain usable images. Researchers intervened as needed, treating the process as iterative because mistakes on live biological samples can waste weeks of preparation and thousands of dollars in materials.
Context-aware operation and image interpretation
During microscopy tests, the system recognized that switching to higher magnification could risk a collision with the sample and adjusted cautiously. It also answered questions about image content, correctly identifying colored structures as cell walls. That combination of instrument control and image interpretation points to a more autonomous laboratory workflow.
Automating live tracking tasks
Researchers then asked the AI to write code to track moving algae in a sample, a task that would otherwise require a scientist to sit for hours and monitor the field of view. The system generated the script, then was pushed to add a user interface so its actions were visible rather than hidden in the background. After revision, the tracker ran successfully for several minutes.
Potential gains for young researchers
Bast argued that such tools could sharply reduce the time needed for graduate students and other researchers to get experimental systems running. Instead of spending two years building and stabilizing a setup, a lab might need only about two months, leaving more time for the underlying biological or physical questions. That shift could change how training and productivity work in experimental science.
Drug discovery as a major target
At Genentech, the same approach is being tested for closed-loop experimentation in pharmaceutical research, where teams may screen thousands, hundreds of thousands, or even millions of molecules. One example involved detecting bubbles in liquid-handling wells, since bubbles can cause inaccurate transfers. The AI adjusted parameters after reading results, reducing bubbles in most wells and showing how faster feedback loops could improve experimental throughput.
CONCLUSION
The effort to give AI direct, standardized control over scientific instruments could make laboratories more automated, safer and faster to iterate. If the approach scales reliably, it may reshape research in fields ranging from biology and drug development to quantum computing and fusion.
Explain this