Programmable science labs allow us to make quantum leap into the AI Science Labs, wherein self-driving labs leverage foundation models to identify the best experiments to run and to run them efficiently in an automated manner in the lab. By combining AI models, feedback loops, and robotics, these labs can execute experiments, analyze results, and refine parameters continuously – without human intervention.
AI Science Labs enable faster discovery, reproducibility, and scalability across scientific domains. They turn traditional, trial-and-error workflows into adaptive, data-driven systems capable of learning from every iteration.
Researchers don’t just run experiments – they reason, execute, and refine continuously. These three correspond to three distinct but connected iterative loops of work. Each loop answers a different question, operates at a different time scale, and optimizes a different objective. Keeping them separate creates clarity, reliability, and explainability – while allowing them to work together as a single system.
Globally, a lot of effort is being invested to build science-specific foundation models to reason over existing scientific knowledge – papers, prior experiments, and domain constraints – to surface promising hypotheses worth testing. Specialized AI models then help identify the most informative experiments to run, designing concrete experimental plans that efficiently validate or refute those hypotheses.
Once an experiment is defined, this loop focuses on efficient iteration. It looks across past runs, parameters, and outcomes to suggest how the next iteration could be improved – adjusting concentrations, timings, scan settings, etc. The goal is faster convergence, higher information gain, and lower cost or risk. Importantly, this loop never changes what is being tested – only how it is executed across iterations.
A specific iteration typically consists of multiple experimental steps. The instrument control loop works operates at the granularity of a single step and focuses on operating the specific instrument as efficiently as possible – while maintaining stability, staying within safety envelopes, and maximizing signal quality. It adapts control policies based on live instrument states but never alters experiment-level plans. This can be visualized as an expert lab scientist operating the instrument – continuously fine-tuning instrument execution while higher-level loops stay focused on science and iteration strategy.
Science progresses by interoperating between the three seamlessly. AI Science Labs, as a result, need to support all three loops. To ensure that the experimental steps can be executed as efficiently as possible, the OSS platform focuses on the inner two loops: the experiment optimization loop and the instrument control loop.
Experiment Optimization Loop
OSS is designed around how real experiments actually progress in the lab – not as one-off runs, but as a sequence of iterations where each result informs what happens next. Researchers constantly adjust experimental parameters as results come in – narrowing ranges, pruning unproductive branches, or stopping early when trends stabilize. OSS captures this process using AI-based experiment optimization models, which are referred to as “world models”.
After each iteration, these world models analyze what parameters were used, what outcomes were observed, and which signals indicate progress or failure. Based on this context, they propose targeted parameter updates for the next iteration – for example, refining a concentration range, adjusting scan resolution, or reallocating effort toward regions showing strong response.
The goal is not to redesign the experiment, but to run it more efficiently: fewer iterations, faster convergence, and better use of time and samples. Over successive iterations, the experiment naturally focuses on the most informative regions of the parameter space.
Instrument Control Loop
Many experimental steps rely on high-end, high-demand instruments where data quality depends heavily on real-time operating conditions. Rather than treating instrument operation as static black-boxes, OSS supports instrument control models that leverage AI-assisted analysis to fine-tune instrument parameters.
For each instrument-specific step, control models continuously monitor instrument signals and measurements as they are acquired. When instability, drift, or suboptimal signal quality is detected, these models can fine-tune control parameters in real time – improving image quality, stabilizing measurements, or reducing retries.
This inner loop operates within a single step, often in seconds or minutes, ensuring that each measurement is as clean and informative as possible before the experiment moves on.
Why This Matters
AI Science Labs require more than intelligent world models and control models – they require infrastructure that cleanly separates intent, execution, and control, while allowing AI to intervene at the right level and time scale.
By combining:
- ULA for precise, auditable experiment definitions
- ULE for deterministic execution and iteration-aware optimization
- ULI for real-world instrument integration and control
OSS provides the universal lab foundation that allows AI Science Labs to operate safely, efficiently, and at scale – across instruments, domains, and lab environments.
Researchers begin by defining a fully specified experiment via ULA: what they want to measure, which instruments are involved, and the constraints they care about (time, sample limits, damage tolerance, and so on). From this, OSS generates a validated experimental workflow that is deterministic, auditable, and safe – the structure of the experiment stays fixed, while parameters can evolve over time.
Once execution begins, experiments proceed in iterations. In each iteration, ULE runs the same logical sequence of steps – often spanning multiple instruments – but with concrete parameter values specific to that iteration. After the iteration completes, OSS captures not just raw measurements, but also derived metrics, quality indicators, and failure signals that summarize how well the experiment performed. With the help of world models, the platform identifies cross-iteration and cross-instrument parameter deltas for refining the next iteration. Each experiment iteration informs the next – reducing redundancy, accelerating convergence, and uncovering new insights faster than manual experimentation ever could.
ULI allows instrument-specific control policies to operate instruments effectively. For this, ULI translates control decisions into concrete actions – either instrument-specific API calls or operator tasks – that dynamically adjust instrument parameters during execution. ULI continuously collects instrument-level signals and logs, allowing control policies to refine measurement parameters in real time and ensure the instrument is used in the most efficient and effective manner.
By integrating structured experiment specifications, AI-driven reasoning, and robotic actuation, OSS enables autonomous feedback loops that continuously refine scientific workflows. OSS, therefore, can be used to build AI-based self-driving labs.