⏱️ Reading time: 8 min
A robot didn’t just choose the right molecule to treat an incurable eye disease: it proposed it, and a real laboratory then confirmed that it worked. Phys.org documented in September 2026 how several groups are now getting their artificial intelligence systems to generate a biological hypothesis and validate it experimentally without a researcher deciding the intermediate step: it’s one of the first public cases of autonomous scientific AI in biology.
📑 En este artículo
The most cited case is Robin, the agent from the FutureHouse lab that in 2025 chained together literature review, hypothesis generation, and experimental design to propose a therapeutic candidate for age-related macular degeneration, and it did so without a human choosing the drug before the trial.
TL;DR
- Robin, from FutureHouse, proposed a drug for macular degeneration in 2025 and validated it in the lab.
- The ROCK inhibitor ripasudil, already approved in Japan for glaucoma, was the candidate Robin suggested repurposing.
- Before Robin, Adam (2009) and Eve (2015) closed that loop in yeast genetics and antimalarials.
- In February 2025, Google introduced its own Gemini-based AI co-scientist for biomedicine.
- The novelty in 2026 is that experimental validation also runs without intermediate human decisions.
What Robin Is
Robin is the orchestration agent that the FutureHouse lab built on top of a family of models specialized in scientific literature, and it embodies the idea of autonomous scientific AI: it chains together literature search, hypothesis generation, and the design of a concrete experiment without a researcher manually choosing the next step.
FutureHouse is a nonprofit lab dedicated to automating science with language models. Before Robin, it already operated specialized agents: one answers questions with citations verified against the literature, another performs deep literature reviews, and another runs data analysis. Robin combines all of them into a single decision chain, from the initial question to the lab protocol.
What Happened
In 2025, Robin received an open-ended question about dry age-related macular degeneration, a disease that progressively damages the retina and for which no approved treatment exists. The agent reviewed the literature on the retinal pigment epithelium (RPE), the layer of cells responsible for clearing photoreceptor debris, and proposed that restoring that cleanup function with a drug already approved for another disease could slow the damage.
The hypothesis pointed to ripasudil, a ROCK enzyme inhibitor approved in Japan for glaucoma. A collaborating lab ran the trial with RPE cells under oxidative stress and confirmed the mechanism Robin had anticipated. No human chose the drug before the experiment: Robin decided which molecule to test and which trial to run, and the human team only carried out the protocol at the bench.
Phys.org documented this type of episode as evidence that autonomous scientific AI systems already exist that are capable of closing their own loop, from question to lab result, without a human arbitrating every intermediate step.
Context and History: From Robot Scientist to Robin
The idea of automating the entire scientific method didn’t start with language models. The robot scientist Adam, built by Ross King’s team at the universities of Aberystwyth and Cambridge, formulated hypotheses in 2009 about which genes coded for orphan enzymes in yeast metabolism, designed experiments to test them, and ran them with its own robotic arm, with no human intervention in that stretch. The work was published in the journal Science.
Its successor, Eve, repeated the scheme in 2015 but aimed at antimalarial drugs: it screened thousands of compounds and found that triclosan, a common antibacterial in toothpaste, also inhibited a key enzyme in the malaria parasite. The finding was published in the Journal of the Royal Society Interface. The difference with Robin is the platform: Adam and Eve ran on fixed, custom-built lab hardware, while Robin orchestrates literature, reasoning, and an external human lab as the hands of the experiment.
In February 2025, Google introduced its own AI co-scientist, a Gemini-based multi-agent system that worked alongside Stanford labs on hypotheses about drug repurposing for liver fibrosis and mechanisms of antibiotic resistance transfer. That system generated ranked hypotheses, but it depended on the human lab deciding which one to test first.
| System | Year | What It Automates | Limitation |
|---|---|---|---|
| Adam | 2009 | Hypothesis and physical execution of the experiment | Fixed hardware, single domain (yeast) |
| Eve | 2015 | Mass compound screening and candidate selection | Requires the original robotic platform |
| Google’s co-scientist | 2025 | Hypothesis generation and ranking | The human lab chooses what to test |
| Robin (FutureHouse) | 2025 | Hypothesis, experiment selection, and result tracking | Setting up the trial still runs in a human lab |
Technical Details: The Mechanism Behind Robin
Robin isn’t a single model, but a chain of specialized agents that hand off work to one another. A first agent searches and summarizes literature with verifiable citations, a second agent turns that summary into concrete, testable hypotheses, and a third agent translates that hypothesis into an executable lab protocol, complete with the reagents and controls a researcher would need to run it.
flowchart LR
A["Research question"] --> B["Literature agent"]
B --> C["Hypothesis agent"]
C --> D["Experimental design agent"]
D --> E[("Human lab")]
E --> F["Trial result"]
F --> CThe critical point in the mechanism is the final jump, from hypothesis to protocol: that’s where Robin chooses, without asking for prior permission, which of several candidate molecules deserves the single trial the human lab is going to run. If that choice is wrong, the entire cycle is wasted, because the lab only has the capacity to test one option at a time.
⚠️ Heads up: no agent of this kind ran the trial with its own hands, someone still mixes the reagents. The autonomy is in the decision, not in the physical execution.
The Race for Autonomous Scientific AI
The Robin episode matters because it reverses the usual order of scientific automation. Until recently, software assisted with the cheap part of the method (searching literature, sorting data) and left the expensive part, deciding which experiment is worth running, in human hands. Robin moves that boundary: it decides the experiment and delegates only the physical execution.
That carries a reputational cost if it goes wrong, and not every attempt to automate science ended well. In 2024, the Sakana AI lab published “The AI Scientist,” a system that wrote and submitted complete papers automatically. The criticism came fast: the texts sounded rigorous, but several of their claims didn’t hold up to careful review. That’s why Robin’s case rests on something different, a result confirmed at a real lab bench, not on a hypothesis simply sounding plausible in a paper.
That distinction, plausible hypothesis versus experimentally confirmed hypothesis, is what separates a useful autonomous scientific AI system from one that only generates convincing text. Even so, the risk of bias is still there: if the literature the agent reads is skewed toward certain mechanisms, the hypothesis inherits that bias before it ever reaches the bench.
What’s Next
FutureHouse has already extended Robin’s family of agents to other biomedical questions, and Google keeps iterating on its co-scientist with new collaborating labs. The bottleneck that remains is the same in both cases: a lab’s physical capacity to run trials, which doesn’t scale at the rate a language model generates new hypotheses.
That leaves two paths open: also automate the lab bench with liquid-handling robotics, as Eve already did in 2015, or accept that the human bottleneck will keep setting the real pace of discovery, no matter how fast the model thinks.
Try it yourself: open FutureHouse’s original publication about Robin and check step by step how many decisions in the scientific process no longer go through a human researcher.
Frequently Asked Questions
What is Robin, FutureHouse’s agent?
It’s an orchestration system that chains together agents specialized in literature, hypothesis generation, and experimental design to propose and validate biological findings with minimal human intervention.
What sets Robin apart from other attempts at autonomous scientific AI?
Unlike Google’s co-scientist, Robin doesn’t just generate and rank hypotheses: it chooses the concrete experiment a human lab must run to confirm them.
What drug did Robin propose for macular degeneration?
It proposed repurposing ripasudil, a ROCK inhibitor approved in Japan for glaucoma, as a possible treatment for dry age-related macular degeneration.
Does Robin run the experiments itself?
No. Robin decides which trial to run, but a human lab still prepares reagents and carries out the physical protocol at the bench.
How is Robin related to the robot scientists Adam and Eve?
They’re the direct precedent: Adam (2009) and Eve (2015) already closed the hypothesis-experiment loop, but on fixed robotic hardware built for that task, not by orchestrating an external human lab.
References
- Phys.org: September 2026 coverage of AI systems that autonomously generate and validate biological discoveries.
- FutureHouse: official site of the lab that developed Robin and its family of scientific agents.
- Wikipedia: Robot scientist: the history of Adam and Eve, Ross King’s robot scientists that closed the hypothesis-experiment loop in 2009 and 2015.
- Google Research Blog: announcements about Google’s Gemini-based AI co-scientist, introduced in February 2025.
📱 Do you like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Featured image: Foto de ThisisEngineering en Unsplash
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments