Robotic liquid handlers play a crucial role in automating laboratory tasks such as sample preparation, high-throughput screening, and assay development. Manually designing protocols takes significant effort, and can result in inefficient protocols and involve human error. We investigate the application of reinforcement learning to automate the protocol design process resulting in reduced human labor and further automation in liquid handling.
We develop a reinforcement learning agent that can automatically output the step-by-step protocol based on the initial state of the deck, reagent types and volumes, and the desired state of the reagents after the protocol is finished. We show that finding the optimal protocol for solving a liquid handler instance is NP-complete, and we present a reinforcement learning algorithm that can solve the planning problem practically for cases with a deck of up to 20 × 20 wells and four different types of reagents. We design and implement an actor-critic approach, and we train our agent using the Impala algorithm. Our findings demonstrate that reinforcement learning can be used to automatically program liquid handler robotic arms, enabling more precise and efficient planning for the liquid handler and laboratory automation.