Self-Driving Laboratories
A self-driving laboratory (SDL) combines automated instruments with closed-loop machine learning: a model of the experiment is updated as data arrive and is used to choose the next experiment, so that informative data are gathered with fewer runs. The basic structure is an iterative loop of design, measurement and analysis; strategies such as design of experiments, active learning and Bayesian optimization differ mainly in how the next experiment is selected and when the loop stops. Building and operating full SDLs is costly and demands skills in domain science, programming and data science, which has motivated small, low-cost instruments for teaching and for prototyping algorithms.
Claude-Light as a low-cost prototyping instrument (2025)
Kitchin introduced Claude-Light, a remotely accessible RGB LED and ten-channel photometer controlled by a Raspberry Pi behind a REST API, as a real but noisy instrument for practicing automation (Kitchin 2025). Using it, he illustrated a hierarchy running from HTTP calls and loops, through a Latin square design, to a Gaussian process model inverted by optimization, and then to LLM-driven instrument selection, tool calling and code generation. The device ran more than 73,000 experiments for more than 1800 unique IP addresses from about August 2024, including classes of about 50 simultaneous students. The paper stresses that reliable uncertainty estimates (helped by output scaling and a white-noise kernel) are essential when uncertainty drives the choice of experiments, and that a single instrument is still far from a lab of interacting instruments. Summary: kitchin2025b.
Claude-Light data for generative inverse design (2025)
Kitchin used 100 Latin hypercube measurements from Claude-Light to compare ways of choosing LED settings that hit target sensor counts, showing that generative models conditioned on the desired outputs recover the same settings as forward-model optimization or a backward model, with sample-based uncertainty (Kitchin 2025). He presents the instrument data as a prototype for formulation-style problems with several inputs and several properties, and notes that how to choose experiments efficiently for training generative models, unlike for forward models in active learning, remains open. Summary: kitchin2025a.
Autonomous laboratories in the agentic science perspective (2025)
In a Comment on agentic science, Xin, Kitchin and Kulik describe a continuum of laboratory automation from robotics to LLM-driven agents (Xin et al. 2025). They cite Coscientist, which turns natural-language requests into procedures and runs cloud laboratory equipment through APIs, and A-Lab, an autonomous solid-state synthesis lab combining robotics, conventional ML and ab initio calculations; they classify A-Lab as not agentic because its LLM only predicts synthesis protocols outside a closed reasoning loop. They argue that autonomous experimentation needs safety interlocks, such as halting a plan that calls for a nonexistent compound or violates safety constraints, and note AFMBench as an early benchmark that exposed failure modes of LLM agents in real microscopy tasks. Summary: xin2025.
Accessibility, deployment models and roadblocks (2025)
In a Perspective with many SDL developers, Canty et al. define an SDL as digital tools that predict, propose experiments and update beliefs coupled to automated hardware that carries them out, and argue that what separates SDLs from high-throughput or cloud labs is the selection of experiments, adaptation of methods and integration of tools into workflows (Canty et al. 2025). They compare centralized SDL facilities with distributed networks of group-level SDLs and favor hybrids of the two (see democratization of research). They list roadblocks including scarce, poorly documented vendor APIs, fragmented middleware (ROS, SiLA 2, BlueSky) adopted mainly by the most advanced groups, the need to encode constraints and make long-term objectives measurable, and the need to measure experimental uncertainties directly. They recommend benchmarking and honest reporting of platform performance, and a consortium and institute to maintain standards, software and training. Summary: canty2025.
Human–AI–robot collaboration in catalysis SDLs (2025)
Orouji et al. frame catalysis SDLs as three components (high-throughput automation, data infrastructure with virtual screening, and AI decision-making) embedded in human oversight, and argue that real practice is a spectrum of semi-autonomous workflows rather than manual versus fully autonomous (Orouji et al. 2025). They propose three nested validation steps before any discovery claim (physics-informed in silico cross-checks, automated replicates to measure noise, and a manual audit), and review examples spanning heterogeneous catalysis (BO-driven fixed-bed screening of 144 CO2 hydrogenation catalysts in 42 days), homogeneous catalysis (Fast-Cat mapping a hydroformylation Pareto front in 45 experiments), LLM-driven platforms (Coscientist, LLM-RDF) and biocatalysis (SAMPLE). They identify catalysis-specific barriers, notably powder handling, high pressure and temperature, and operando probes incompatible with robots, and name hybrid computational–experimental workflows, proxy measurements, synthesis–testing co-design and open science as opportunities. Summary: orouji2025.
Self-driving models as the computational counterpart (2026)
The self-driving lab paradigm is extended to computation with “self-driving models”, which automate model building and refinement against multimodal catalysis data rather than synthesis and testing (Medford et al. 2026). The authors suggest such models could eventually be embedded in self-driving laboratories as an advanced form of model-based experimental design or “knowledge engine”, and that their outputs remain interpretable because they are explicit physical or statistical models. Summary: medford2026.
Stages of autonomy in catalysis laboratories (2026)
A Nature Catalysis roadmap describes autonomous catalysis laboratories as developing through five stages: task automation, workflow automation, intelligent planning, closed-loop automation, and AI-empowered human-machine collaboration (Xin et al. 2026). It cites a Bayesian optimization platform that synthesized and tested 144 CO2-to-methanol catalyst formulations in six weeks, but notes that current closed-loop systems remain tool-centric, running human-designed protocols rather than setting their own goals. Robotic hardware is not yet versatile enough and orchestration software is fragmented, so the authors expect a gradual continuum of increasingly capable laboratories.
Their longer-term proposal is a network of cloud laboratories that share abstractions while using different instruments. A planner-executor checks actions against a digital twin that includes deactivation and regeneration, uses Bayesian optimization for single-step choices and model-based or offline reinforcement learning for multi-step sequences, and works within a guardrail stack (quality-control runs, drift monitoring, human review of high-uncertainty proposals, replication and shadow-mode testing). Summary: xin2026.
Autonomy levels and safety reporting (2026)
The TRACE-AI guidelines classify closed-loop experimentation into autonomy levels A0 (manual) to A5 (distributed autonomy across laboratories), by analogy to autonomous driving, and ask authors to declare the level and any human interventions, including manual changes to the exploration and exploitation balance (Xin et al. 2026). They also call for safety sandboxes with human-set parameter bounds, layered interlocks that shut down on threshold exceedances, leaks or loss of communication with the agent, timestamped logs, and stopping criteria specified in advance. Summary: xin2026a.