Robot Scientists & Lab Automation

This page describes the Robot Scientists & Lab Automation row of the Papers.md matrix: systems that couple AI to physical laboratory automation, closing the loop from hypothesis to robotically executed experiment. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The discriminator is physical or automated wet-lab execution. An in-silico agent alone does not qualify, however autonomous its reasoning, which is what separates this row from Scientific Literature & Discovery Agents and the two biomedical-agent rows. A paper belongs here when hypotheses reach a robot, a bioreactor array or an instrument, and results come back.

That makes this the most cell-ag-relevant of the agent rows by some distance, and the one where a paper’s claims are hardest to overstate: an autonomous system that runs cell culture has to survive contamination, drift and hardware failure, not just an evaluation set. It is also the oldest, running from #182 in 2004 through the current LLM-driven systems, so it is the one row where the pre-LLM and post-LLM approaches can be compared directly.

Cellular Engineering

#16 (Kanda et al. 2022, eLife, RIKEN): the closest existing analogue to an autonomous cultivated-meat differentiation campaign. A LabDroid robot coupled to a batch Bayesian optimization algorithm autonomously searched the induced differentiation of iPSC-derived retinal pigment epithelial cells, from a space of 200 million possible parameter combinations, running 143 conditions over 111 days and ending with 88% better iPSC-RPE production than the pre-optimized protocol on pigmentation score. The framing in the abstract is the part worth carrying: induced differentiation is among the most experience- and skill-dependent processes in the field, and establishing optimal conditions often takes years, which is exactly the bottleneck a cultivated-meat programme faces when moving a new species or cell line into production. Also in Bayesian Optimization. Code at labauto/LabDroid_optimizer.
#66 Lila (Singh et al. 2023, bioRxiv, Amyris): an automated scientist covering the whole design-build-test-learn cycle for microbial strain engineering, notable here for how much of the “AI” is classical and explicit. Its Route Finder runs flux-balance analysis with COBRApy over genome-scale models of S. cerevisiae and E. coli extended with a universal set of heterologous reactions, using an objective weight vector that penalizes heterologous reactions roughly 100-fold more than native ones so the minimal foreign pathway is found. Its enzyme-selection algorithm ranks candidate sequences by five criteria in fixed priority order, beginning with the fraction of PubMed abstracts co-mentioning the enzyme and the target molecule or host, then BRENDA kinetic constants, UniProt protein or mRNA evidence, phylogenetic topology and alignment diversity, tightening or relaxing thresholds iteratively until the candidate count is synthesizable (often 100 candidates to four within ten iterations, under five seconds). Those criteria came from about a dozen hours of interviews with human strain engineers. Lila attempted 454 molecules and produced strains overproducing 242, of which 180 are non-native, including the highest published titers for naringenin. Humans stay in the loop to curate the metabolic model and suggest protein modifications.

Bioprocess & Scale-Up

#61 Agentic Lab (Wang et al. 2025, bioRxiv): a human-in-the-loop system rather than a robot, and deliberately so. A virtual PI agent, MolAgent, orchestrates specialized subagents for knowledge retrieval, protocol design, multimodal analysis and training-free segmentation, built on LangGraph with an agent-as-tool interface and a ReAct loop, using unmodified multimodal models from several vendors to preserve general adaptability. The physical coupling comes through an augmented-reality interface: the system perceives what the human operator is actually doing, gives context-aware instructions, and flags procedural errors in real time for correction. Its retrieval subagent generates five to eight strategic subqueries across protocols.io, PubMed, FPbase and Scholar, scores sources by reliability, date, length and relevance, and returns a cited knowledge report. Demonstrated on organoid differentiation from human pluripotent stem cells, where it generated protocols, monitored culture and identified morphological heterogeneity linked to growth conditions. Also in Agent Infrastructure.

AI Tooling / Methodology

#182 The Robot Scientist (King et al. 2004, Nature): the founding paper of the row and still the cleanest statement of what a closed loop is. A physically implemented robot originated hypotheses, designed experiments, ran them, falsified hypotheses inconsistent with the data, and repeated, applied to yeast gene function via auxotrophic growth of deletion mutants against a Prolog model of the aromatic amino acid synthesis pathway translated from KEGG. The economics are the memorable result and the reason the paper still matters: metabolite costs spanned roughly three orders of magnitude, from anthranilate at a normalized cost of 10 to phosphoenolpyruvate at 9,385, and an intelligent experiment-selection strategy was competitive with human performance while cutting cost 3-fold against cheapest-first selection and 100-fold against random. Any cell-ag team choosing which media component to test next is solving the same problem. Also in Active Learning.
#64 Genesis (Tiukova et al. 2024, arXiv, King group): the third generation after Adam (the first machine to autonomously discover novel scientific knowledge) and Eve (early-stage drug design, whose most cited result is identifying triclosan as a DHFR inhibitor in malaria parasites). Genesis targets automatic improvement of systems-biology models with thousands of interacting causal components, and its stated design point is one thousand computer-controlled micro-bioreactors running up to one thousand hypothesis-led closed-loop cycles per day. Supporting infrastructure includes AutonoMS for high-throughput mass spectrometry, Genesis-DB for agent-accessible structured domain data, RIMBO to describe model revisions, and LGEM+ for abductive improvement of genome-scale metabolic models. The paper’s argument for why automation is necessary is worth reading in a cell-ag context: current high-throughput methods generate volume but are not hypothesis-led, so each individual experiment is not designed to test anything.
#65 AutonoMS (Brunnsåker et al. 2025, bioRxiv, King group): couples LLM agents to automated cell culture and metabolomics, with a logical scaffold in between because, as the paper puts it, LLMs struggle to interact with logical structures. Hypotheses are generated by relational learning, frequent pattern mining with the ILP engine Aleph over a Datalog database built from SGD, ChEBI, STITCH-DB and the Yeast9 consensus model, restricted to observables the lab can actually act on. Patterns are propositionalized across 4,678 single-gene deletants and weighted by an ElasticNet predicting amino-acid accumulation, so each hypothesis carries a testable association rather than a bare assertion. Validated in S. cerevisiae, finding glutamate-induced synergistic growth inhibition in spermine-treated cells and aminoadipate’s partial rescue of formic-acid stress. All hypotheses, experiments and data are recorded in a graph database using controlled vocabularies, with hypotheses represented in description logics. Code at DanielBrunnsaker/GenExp; data at Zenodo.
#62 BioMARS (Qiu et al. 2025, arXiv): the fully robotic counterpart to #61, with three agents in a hierarchy: a Biologist Agent that synthesizes protocols by retrieval-augmented generation with a Knowledge Checker and Workflow Checker, a Technician Agent that translates them into robotic pseudo-code through a Code Generator and Code Checker, and an Inspector Agent that monitors execution on two levels, a vision encoder detecting spatial anomalies against keyframes and a VLM reading the scene for high-level procedural inconsistencies. The hardware is a dual-arm collaborative robot with four modular workstations, automated centrifuge and CO₂ incubator, coordinated over ROS with a standardized API of 11 base functions so new instruments plug in. It autonomously performs cell passaging and culture, matching or exceeding manual work on viability, consistency and morphology, and beat conventional strategies at optimizing RPE differentiation. The checker-at-every-stage design is the notable part: three of its components exist only to catch the previous component’s errors. Code at AlexandreQ27/BioMARS.
#154 Robin (Ghareeb et al. 2026, Nature, FutureHouse): described in full under Scientific Literature & Discovery Agents. It is in this row because the loop reaches the bench: proposed compounds were ordered and assayed, ripasudil and KL001 confirmed in vitro, and a follow-up RNA-seq experiment proposed and analysed.
#274 ORGANA (Darvish et al. 2025, Matter): a robotic assistant for chemistry, included here rather than in Chemistry / Synthesis Agents because its contribution is the physical execution layer. It perceives the scene and estimates object poses, asks the chemist to ground which beaker is which (necessary when several objects of one type are present), derives goals through an LLM dialogue that handles ambiguity, then plans a parallel schedule minimizing total time and executes it across instruments asynchronously. It compares results against the reasoner’s expectations mid-run and escalates to the user only on inconsistency. Demonstrated on solubility, pH, recrystallization and electrochemistry, including a 19-step parallel plan characterizing quinone derivatives for flow batteries, where its measured Pourbaix slopes (−61.3, −61.8, −61.0 mV/pH) match the theoretical −59. Its user study is the unusual part and the most transferable: over 50% reductions in reported frustration and physical demand, and an average 80.3% time saving. Code at ac-rad/organa.

The row splits cleanly on one question, which is where the human sits. #182, #64 and #62 aim to remove the human from the loop; #61, #66, #274 and #154 keep them in it deliberately, for curation, grounding, or physical execution. Both point at the same destination and the papers differ mainly in how much oversight they retain, which is the choice a cell-ag facility would be making rather than a difference in ambition.

Adjacent methods

Bayesian Optimization and Active Learning: the experiment-selection machinery that decides what the robot runs next, and where #16 and #182 are also placed.
Scientific Literature & Discovery Agents: the hypothesis-generation half of these loops.
General-Purpose and Domain-Specific Biomedical Agents: the in-silico agents that stop short of physical execution.
Chemistry / Synthesis Agents: Coscientist’s EXPERIMENT command reaches the same hardware layer from the chemistry side.
Agent Infrastructure (Frameworks, KGs, Protocols): the orchestration and knowledge layers underneath.