When people say they want to automate pipetting, they usually mean one of two very different things, and the confusion between them causes a lot of wasted effort. One is the physical act: making a machine perform a single transfer as well as a trained hand does. The other is the choreography: sequencing hundreds or thousands of transfers so the whole run finishes quickly and correctly. Two recent papers tackle these separately, and reading them side by side is the clearest way to see that automation is not one problem but two, each with its own kind of difficulty. This is the second in our series dissecting research for people who run the deck.
The act: a robot holding a human pipette
The first paper is Zhang, Wan, Tanaka, Fujita, and Harada, Integrating a Manual Pipette into a Collaborative Robot Manipulator for Flexible Liquid Dispensing (arXiv:2207.01214, cs.RO, submitted July 2022). Instead of buying a dedicated liquid handler, the authors give a six-degree-of-freedom collaborative robot an ordinary handheld pipette and teach it to use the tool the way a person would. The system has three parts: a custom end-effector that grips the pipette body and presses the plunger, computer vision that recognizes where the labware actually sits, and visual classifiers that catch and correct placement errors before they ruin a transfer.
What makes this interesting is not that a robot can pipette, it is the choice to keep the human tool. A manual pipette encodes decades of design for accuracy, the plunger stops, the tip geometry, the ergonomics of a clean dispense, and by driving that tool rather than replacing it, the robot inherits all of that calibration for free. The trade the authors are explicit about is throughput: this approach suits low-frequency, repetitive work rather than high-speed screening, because a robot arm reaching for a benchtop pipette will never match a purpose-built head firing eight or ninety-six channels at once. It is automation for the lab that has the pipette and the protocol but not the budget or the volume to justify a dedicated instrument, and that is a large fraction of labs.
For the themes we care about, the lesson is that the hard part of the physical act is not the plunger, it is everything around it: knowing where the plate is, confirming the tip seated, recovering when something is a millimeter off. Those are exactly the tacit skills a bench technician performs without thinking, and the paper is a catalogue of what it takes to make them explicit enough for a machine.
The choreography: a run as a routing problem
The second paper is Wu, Wang, and Coley, Optimization of Robotic Liquid Handling as a Capacitated Vehicle Routing Problem (arXiv:2506.02795, math.OC and cs.RO, submitted June 2025). It ignores the individual transfer entirely and asks a different question: given a whole plate map of sources and destinations and a multichannel head that can carry only so much at once, what is the fastest order to do everything? The insight is that this is structurally the same problem as routing delivery trucks, the capacitated vehicle routing problem, a classic in operations research with a deep toolbox of heuristics ready to borrow.
The results are the kind that make an operations person sit up. On synthetic tasks the method cut execution time by about 37 percent, and on a real high-throughput materials-discovery workflow it saved 61 minutes of instrument time after just 3 minutes of computation, with no hardware changes at all. That last part is the point: the same robot, the same tips, the same liquids, simply visited in a smarter order. The authors are careful to frame it as a heuristic strategy rather than a guaranteed optimum, which is honest and also exactly why it is practical, since a good-enough route found in minutes beats a perfect route that takes longer to compute than to execute.
Why the pair matters together
Put the two papers next to each other and the shape of automation comes into focus. The first is about making one transfer trustworthy; the second is about making a thousand transfers efficient. They barely overlap, which is precisely the message. A lab that nails the act but ignores the choreography ends up with beautiful transfers and a run that takes all afternoon. A lab that optimizes the choreography on top of transfers it has not made reliable just reaches the wrong answer faster. The two problems have different owners, different tools, and different failure modes, and treating them as one thing, "automation," is how projects stall.
There is a deeper point for anyone who thinks in liquid classes. The routing paper optimizes time while holding the transfers themselves fixed, which only works if each transfer is a known, reliable quantity, if the liquid class is trustworthy. The efficiency layer sits on top of the correctness layer and depends on it completely. Get the class right first, in the sense the first paper cares about, and only then does the choreography the second paper optimizes have solid ground to stand on.
Automation is two problems wearing one word. One paper makes a single transfer worthy of trust; the other makes a thousand transfers worthy of the clock. You need both, in that order.
References
- J. Zhang, W. Wan, N. Tanaka, M. Fujita, K. Harada. Integrating a Manual Pipette into a Collaborative Robot Manipulator for Flexible Liquid Dispensing. arXiv:2207.01214 (cs.RO), 2022. arxiv.org/abs/2207.01214
- G. Wu, R. Wang, C. W. Coley. Optimization of Robotic Liquid Handling as a Capacitated Vehicle Routing Problem. arXiv:2506.02795 (math.OC, cs.RO), 2025. arxiv.org/abs/2506.02795