Nanoliter Droplet Dispensing via Active Inference
This work was part of my master’s thesis, Vision-Based Control of Complex Physical Systems via Active Inference.
Introduction
- Motivation: Combine active inference with vision-based measurements and physical process knowledge in the perception and learning components of an active inference controller.
- Application: Calibration of a nanoliter droplet dispenser. The objective is to minimize the number of droplets used during calibration, since each droplet increases production cost.
Methods
- Implementation of an Active Inference Controller (AIC) using variational message passing on a Bayesian network implemented with RxInfer (see Figure 1).
- The Bayesian network connects: vision-based droplet measurements, a physical model of the dispenser and the next control action.
- Performance metric: Sample efficiency (number of droplets required for calibration).
- Baselines: Comparison with Bayesian Optimization (BOC), the state-of-the-art method in the literature, and PID control, the industrial standard.
Figure 1: Schematic of the implemented control loop.
Results
Figure 2: Summary of the results:
(a) Distribution of sample efficiencies over 10,000 simulations,
(b) Comparison of BOC and AIC in an example dispensing simulation,
(c) Average sample efficiencies of the controllers shown in (a).
The active inference controller improves sample efficiency by 85% compared to BOC and 92% compared to PID control.
Discussion
The active inference approach provides a more effective exploration–exploitation trade-off for calibration tasks than optimization-based methods and classical feedback controllers.
Conclusion
An active inference-based controller was developed for calibrating a nanoliter droplet dispenser. Simulation results demonstrate a substantial improvement in sample efficiency, reducing the number of required droplets by 85% compared to the state-of-the-art Bayesian optimization controller.