There is a durable joke in drug discovery, which is that the industry spends 10 to 15 years and something north of $2 billion to bring a single drug to market, and much of that time is spent doing the same manual bench work over and over while an enormous amount of the data - the failed runs, the botched pipetting, the timing that was slightly off - gets thrown directly in the trash. This is, if you think about it, a strange way to run an information business.
Medra's argument is that biology is, secretly, a robotics problem. Not a modeling problem, not a compute problem - a robotics problem. You can build an AI that proposes a thousand experiments, but if a human has to run all thousand by hand, you have not actually sped anything up. You have just made a very well-read intern who cannot use their hands.
So Medra built the hands. The company calls the result a Physical AI Scientist: a system that unifies robotics, reasoning and data generation into one continuous loop. It designs an experiment, a robot executes it, cameras and sensors watch what happened, the result feeds back into the model, and it tries again - overnight, on weekends, without asking for a coffee break.
The founder is Michelle Lee, who did her PhD at the Stanford AI Lab on physical AI and reinforcement learning, and who had, by 2022, done the hard part of an academic career: she had an assistant professorship at NYU lined up. She left it. The bet was that the interesting version of robotics was not going to happen inside a university lab, and that the place it would happen was the wet bench, where the data is hardest to fake and hardest to get.
What makes this more than a pitch deck is that the robots do not use custom hardware for every task. They have learned to operate more than 75% of standard laboratory instruments - the same machines a human scientist already owns - and they are told what to do in plain English, roughly the way you would text instructions to a lab mate. The technical name Medra uses is a vision-language-lab-action model, which is a mouthful, but the idea is simple: the machine watches, it listens, and it moves.
The commercial discipline is worth noting too. Medra does not do free pilots. Its early customers are paying biopharma partners - among them Genentech, the Roche subsidiary, and Addition Therapeutics - which is a useful filter, because the people who invented modern biotech are not easily impressed by a demo. As of late 2025 there were five units running at partner sites across the US, and the company had raised about $63 million to build one of the largest autonomous labs in the country.
