Cem Seref Toker
← All essays

Why Robotics Will Revolutionise Scientific Discovery

Building the infrastructure for a future where science never stops.

We have made extraordinary scientific progress. But the everyday work of discovery is still slow. Scientists spend years developing hypotheses, designing experiments, preparing samples, operating equipment, and analysing results. Then they do it all again.

AI is changing how quickly we can explore ideas in biology, chemistry, and physics. Models can predict molecular structures, suggest new compounds, and help generate hypotheses. Yet every promising idea eventually meets the same test: does it work in the real world?

AI can generate hypotheses much faster than we can experimentally validate them. Coming up with ideas is only part of the problem. We also need a way to test them.

The experimental bottleneck

FIG. 01The experimental bottleneck
Many ideasPhysical testingNew evidence
Ideas are abundant. Testing them is the constraint.

Data is one of the biggest bottlenecks in scientific discovery. Language models can learn from enormous amounts of information that already exist online. Scientific AI also needs something much harder to obtain: experimental evidence about parts of the physical world we have barely explored.

We do not know how most possible molecules interact with biological systems, how cells respond to countless combinations of conditions, or which unexplored materials might have useful properties. We cannot simply scrape these answers from the internet. We have to run experiments.

Today, that means expensive equipment, skilled researchers, and a great deal of hands-on work. The constraint is not just how much AI we can train. It is how much new, reliable scientific data we can generate.

Making science programmable

At Neuromorphic, we believe physical experimentation should become as programmable and scalable as computation.

Imagine a laboratory where robots prepare samples, move materials between instruments, operate equipment, and run experiments throughout the day and night. A scientist would describe an experimental objective, and an intelligent system would translate it into a sequence of physical actions.

FIG. 02A connected laboratory
PlanExecuteMeasure
Robotics connects the plan, the instruments, and the result.

This does not mean replacing every instrument in the lab. General-purpose robotics can connect the equipment scientists already use, turning separate instruments into a flexible system.

Instead of manually coordinating every step, researchers could work with laboratories capable of running thousands of experiments with consistency, traceability, and far less manual intervention. The lab becomes an experimental engine.

Closing the loop

The most interesting part is what happens after an experiment finishes.

An AI system could analyse the result, identify what remains uncertain, and help choose the next experiment. Rather than simply working through a fixed list, the laboratory could adapt as it learns.

FIG. 03A laboratory that learns
BetterhypothesesBetterexperimentsBetterevidenceMore knowledgewith every cycle
The output of one experiment becomes the starting point for the next.

Every experiment adds evidence. That evidence helps improve the models, which can then propose better experiments. Better experiments, in turn, produce more useful data.

The goal is not simply to run more experiments. It is to learn more per unit of time.

That distinction matters. A laboratory that runs quickly is useful. A laboratory that gets better at deciding what to investigate could change how discovery happens. Over time, this could accelerate work in drug development, cancer research, materials science, and many other fields.

The future: biological data centres

Data centres changed computing by making computational power available on demand. I think laboratories could undergo a similar shift, making experimental capacity available whenever a scientist needs it.

Think of networks of robots, specialised instruments, and scientific AI working together. These systems could investigate questions, generate reliable experimental data, and use the results to guide what comes next. Discovery could become a continuous process, rather than a series of experiments limited by how much manual work a team can fit into a day.

Scientists would spend more of their time choosing important problems, developing theories, and interpreting discoveries. Autonomous laboratories would take on more of the work of executing and refining experiments at scale.

This is the future we want to help build at Neuromorphic: physical infrastructure that lets AI move beyond learning what humanity already knows and start helping us discover what we do not.