Abstract
High-throughput automated science relies heavily on robotic liquid handlers, yet these systems often execute transfer tasks in inefficient sequences. This study demonstrates that by modeling liquid handling as a Capacitated Vehicle Routing Problem (CVRP), robotic execution time can be reduced by up to 37% without hardware modifications. By treating an 8-channel pipette as a delivery vehicle with a fixed capacity, the authors utilize off-the-shelf optimization solvers to significantly enhance laboratory throughput.
Introduction
In modern automated laboratories, high-throughput combinatorial screening is essential for discovering new drug combinations and advanced materials. These experiments involve transferring thousands of distinct liquid solutions from source plates to destination plates. While downstream analysis is often rapid, the physical liquid handling remains a significant bottleneck. A standard 350-transfer task can take over an hour, largely because control software often executes transfers in naive, unoptimized sequences (e.g., row-by-row).
An 8-channel pipette can theoretically transfer up to eight liquids simultaneously, but only if the target wells are physically adjacent. Inefficient sequencing leads to the robotic arm moving back and forth to pick up single samples, wasting time and resources. The authors set out to find a systematic, algorithmic way to optimize the sequence of these pipetting jobs to minimize total execution time.
Methodology
The authors address this inefficiency by framing the pipetting schedule as a Capacitated Vehicle Routing Problem (CVRP), a classic operations research challenge used by logistics companies like FedEx. In this analogy, the 8-channel pipette acts as a delivery truck with a capacity of 8 deliveries per trip. Each required liquid transfer is a location to be visited, and the "distance" between jobs is defined by the physical geometry of the well plate. Adjacent wells in the same row have an effective distance of zero, while non-adjacent wells require a full robotic arm movement (distance of 1).
By defining a mathematical proxy for execution time—incorporating aspiration speed, volume, and tip movement—the authors utilized Google OR-Tools to generate highly parallelized work lists. This approach proves that smarter software can substitute for faster hardware. The mathematical proxy for execution time correlated almost perfectly ($R^2 > 0.99$) with actual simulated execution times, validating the simplicity of the model.
Experimental Setup and Results
The approach was evaluated on synthetic datasets across standard labware formats (96- to 1536-well plates) and a real-world benchmark from an autonomous materials discovery campaign aimed at finding polymer blends that stabilize enzymes.
- Synthetic Performance: The CVRP method achieved an average 15% reduction in execution time compared to "long-axis prioritized" (LAP) heuristics and up to 37% reduction over baseline row-major sorting.
- Real-World Impact: In a polymer blending campaign, the solver saved 61 minutes of physical robot time with only 3 minutes of computation, representing a 25% throughput improvement. This is particularly significant as AI-driven experiments often create highly non-uniform, clustered transfer patterns that naive heuristics struggle to handle.
- Scalability: For a massive task of 14,000 transfers on a 1536-well plate, the solver computed a route in 120 seconds that saved 158 minutes of execution time.
- Platform Independence: The optimized schedules were successfully validated on both Tecan Evo 200 and JANUS G3 workstations, proving the software is hardware-agnostic.
Discussion
This work demonstrates that repurposing logistics algorithms for lab automation is a high-impact, low-cost strategy. The CVRP solver dramatically outperforms traditional heuristics, especially in dynamic workflows where transfer patterns become highly skewed over time.
While the CVRP solver uses heuristics for this NP-hard problem and requires a minimum compute budget, its performance is robust. Future work may involve moving further up the pipeline: instead of just optimizing the execution of a pre-defined list, the experimental design algorithm itself could be made "layout-aware." By intentionally assigning destination wells in a way that favors the geometry of the 8-channel pipette, researchers could reduce robotic overhead even further.
Summary updated