Self-Driving Lab
A self-driving lab, also called a self-steering laboratory, is a research platform that plans, runs and analyses experiments with minimal human intervention. Robotic systems carry out the experiments, and algorithms use the results to select the next experiment. Researchers set the goal and constraints. The system then runs through the experimental cycle largely without human intervention.
How does a self-driving lab work?
A self-driving lab operates in a closed loop with four stages: experimental design, execution, analysis and adjustment. An algorithm selects a material variant or a process condition to test, and a robot runs the experiment. The resulting samples are then automatically characterised and the data analysed. The system uses these results to plan the next experiment. The cycle continues without a person having to initiate each step.
This allows researchers to test far more material variants in less time than in a traditional, manually run lab.
Automated laboratories and self-driving labs: What is the difference?
Not every automated laboratory is a self-driving lab. Automation can simply mean using robots for individual tasks, such as preparing samples or taking measurements. A researcher still decides which experiment comes next. For example, they might use design of experiments (DoE) to define 50 combinations for a robot to test in sequence.
The full set of experiments and their sequence are fixed before testing begins.
A self-driving lab goes one step further: it also automates decisions about what to test next. After each experiment, an algorithm analyses the results and uses them to choose the next experiment, instead of working through a fixed list. A self-driving lab therefore relies on automation for both running experiments and deciding which ones to run.
Why data structure remains a prerequisite
The quality of a self-driving lab’s decisions depends on the data it uses. Material compositions, process parameters and measurement results must be consistently structured, clearly linked and machine-readable so that an algorithm can interpret them correctly.
Without this structure, even advanced AI models cannot make reliable decisions, and robotic systems cannot run experiments reliably. This can happen when formulation and measurement data are scattered across separate systems or stored in inconsistent formats. Structured data is therefore essential to a self-driving lab, however much of the physical work has already been automated.
Opportunities and limitations
Self-driving labs can test far more material variants and shorten development cycles. They are particularly useful when there are many combinations of variables to explore, for example in the case of functional materials for energy applications.
However, setting up a self-driving lab requires substantial investment in robotics, sensors and software integration. Workflows involving complex or rarely studied classes of materials cannot always be fully automated. Even a well-developed system still needs structured, reliable data to make sound decisions.
Application example from research
The Energy Materials Acceleration Platform (E-MAP) at the Karlsruhe Institute of Technology provides one example. Robots prepare materials, handle samples, and deposit and characterise thin films, among other tasks. E-MAP is considered a self-driving lab — a largely autonomous research platform of the kind supported by Germany’s HighTech Agenda.
The platform illustrates how investment in robotics and software integration supports everyday research. It also shows that robots alone are not enough. Linking characterisation with data analysis is essential for researchers to make greater use of data in guiding experiments.
FAQ answers: Self-Driving Lab
What is a self-driving lab?
A self-driving lab plans, runs and analyses experiments in a closed loop with little user intervention between cycles. An AI model proposes the next experiment based on previous results. Robots carry out the experiment, and the measurement data is automatically fed back into the model to guide its next proposal. This cycle of prediction, experiment and feedback repeats until the development goal is reached or a predefined stopping condition is met.
In traditional automated experimentation, a researcher defines the entire series of experiments in advance. In a self-driving lab, the model uses the results of each cycle to choose the next experiment with the greatest potential to provide new insights.
How does it differ from an automated laboratory?
An automated lab runs a predefined series of experiments. The full set of experiments and their sequence are fixed before testing begins.
A self-driving lab goes one step further: experiment planning is itself part of the automated cycle. After each experiment, the model analyses the results and uses them to choose what to test next, rather than following a fixed list.
This distinction matters because the two approaches can overlap in practice. An automated lab that uses Bayesian optimisation can function much like a self-driving lab, even if robots do not carry out every physical task. Bayesian optimisation is also used in predictive AI applications for formulation development.
What role does data structure play?
It is essential. To choose the next experiment independently, a model must be able to interpret the previous cycle’s results immediately and correctly, without anyone having to prepare the data manually. Without structured data, this manual preparation becomes a bottleneck in a step that should be fully automated.
In practice, measurements must follow a consistent format, and materials and formulations must be clearly identified. Links between process parameters, material properties and test results need to be recorded when the data is collected. A self-driving lab’s speed and reliability therefore depend on the data structure supporting each experimental cycle. Unstructured or scattered data slows down the entire loop, not just the analysis, because every cycle relies on well-organised, reliable data.
What are the limitations?
A self-driving lab optimises within predefined limits. The development team decides which properties matter, which variables can change, and what regulatory or economic constraints apply. The system works within those limits; it does not set them.
There are practical limits to automation, too. Some experimental setups cannot be automated using robots. For complex tests or small numbers of samples, the effort involved may outweigh the benefits.
Rare or unfamiliar effects also pose a challenge when little or no training data is available. The model’s recommendations become less certain, so the system needs to focus more on targeted exploration: testing conditions whose outcomes are uncertain but could provide useful information, rather than rapidly improving approaches that already show promise.
Finally, predictions and experiment proposals do not explain why a result occurred. Researchers still need to determine whether an unexpected result stems from a measurement error, a scale-up effect or a previously unknown interaction.