WetRobo: A Reproducible Robot Kit for Coding Agents in Biological Laboratories
The paper introduces WetRobo, a reproducible robot kit that enables a coding agent to autonomously adapt and execute natural-language biological tasks across different laboratories by writing and executing code, thereby overcoming the transfer limitations of traditional vision-language-action policies.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the quiet hum of a biological laboratory, a great deal of work depends on the steady, repetitive hands of human researchers. They feed cells, move delicate glass dishes, and open heavy doors to incubators, tasks that must be performed with precision day after day. For decades, scientists have hoped to hand these duties over to robots, freeing human experts to focus on discovery rather than manual labor. The most promising recent approach has relied on artificial intelligence systems trained to see the world and act within it, much like a person learning a new skill by watching a video. These systems, often called vision-language-action models, are taught by showing them thousands of examples of a robot arm moving a bottle or lifting a lid. The hope is that once trained, the robot can simply look at a new situation and know what to do. However, in the messy reality of a working lab, where lighting shifts, equipment varies, and objects are placed slightly differently, these trained systems often stumble. A robot that masters a task in one room may fail completely when moved to the next, because the subtle changes confuse the patterns it learned. This fragility has left many laboratories waiting for a more adaptable solution.
A team of researchers has now introduced a different path forward with a system they call WetRobo. Instead of trying to train a robot to memorize a specific set of movements, they built a kit that allows a computer program to write its own instructions on the spot. The kit includes a standard robot arm, common laboratory equipment like an incubator and reagent bottles, and a set of recorded videos showing how a human might perform a task. Crucially, it also includes a guidebook written in plain language that tells a coding agent how to approach the job. This agent is a type of artificial intelligence capable of writing computer code. When a researcher gives the robot a simple command in natural language, such as "lift the bottle cap," the agent looks at the local environment, consults the guidebook, and writes a custom program to solve the problem. It does not rely on a pre-trained brain that might be confused by a change in camera angle or a different brand of gripper. Instead, it observes the scene, measures distances, and adjusts its own code until the task is done.
The researchers tested this approach in two distinct laboratories, which they called Lab X and Lab Y. These rooms were not identical; they had different lighting, different numbers of cameras, and even different types of robot grippers. In one experiment, the robot was asked to remove a screw cap from a reagent bottle. The coding agent, using the WetRobo kit, successfully figured out how to grasp and twist the cap in both laboratories. It did this by writing code that used the cameras to locate the white cap, measured the distance to it, and then adjusted its movements based on what it saw. In Lab Y, where the robot used a standard gripper, the agent wrote a program that stopped if the arm felt too much resistance. In Lab X, where the robot had a different, custom-made gripper that could not measure force in the same way, the agent wrote a different program that relied on visual checks and simulated collisions to avoid errors. The result was a robot that could adapt its own instructions to fit the specific room it was in.
To see how effective this method was, the researchers compared it against the traditional approach of training a vision-based model. They took a system that had been fine-tuned on thousands of video demonstrations from Lab X and asked it to perform the same bottle-cap task. In Lab X, where it had been trained, the system worked perfectly. But when the same system was moved to Lab Y, with its different lighting and equipment, it failed completely. It could not transfer its knowledge to the new setting. The coding agent, by contrast, succeeded in both places. It did not need to be retrained or shown new videos; it simply observed the new room and wrote a new set of instructions that worked. The researchers also tested the system on other tasks, such as lifting the lid off a petri dish and opening the heavy door of an incubator. In each case, the agent observed the local conditions, wrote code to handle the specific geometry of the door or dish, and executed the task successfully.
The process of adaptation was visible in the code the agent produced. When the robot struggled to align its gripper with a bottle cap, the agent would write a small program to test the movement, check the camera image, and then adjust the approach. If the robot slipped, the agent would rewrite the code to include a slower, more careful pull. In one instance, the agent realized that shadows on the incubator door were misleading its vision, so it wrote a new program to measure the door's angle using depth data instead of light. This ability to inspect its own mistakes and rewrite its own logic allowed the robot to overcome challenges that would have stumped a pre-trained system. The researchers noted that the agent often reused parts of the code it had written for previous tasks, such as the logic for grasping a thin object, to solve new problems. This meant that as the robot worked in a lab, it built up a library of useful instructions that could be shared with other labs, making future setups faster and easier.
This work suggests a practical route for bringing robotics into biological research. Rather than requiring every laboratory to collect hours of video data and train expensive neural networks for every new robot, scientists can distribute a kit and let a coding agent adapt it to the local environment. The researchers demonstrated that this method allows a robot to handle real-world variations in equipment and lighting that break traditional systems. While the study was limited to a few specific tasks and a single type of coding agent, the results point to a future where robots can be deployed in diverse settings without the need for extensive retraining. The key finding is that giving a robot the ability to write its own instructions based on what it sees is a more robust way to handle the unpredictable nature of a working laboratory than trying to teach it a fixed set of movements. By shifting the burden of adaptation from human trainers to the software itself, WetRobo offers a way to make laboratory automation more flexible, accessible, and ready for the real world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.