On-Device Reinforcement Learning Controller Manages Automated Insulin Delivery in Simulation Tests

Automated insulin delivery systems integrate continuous glucose monitoring with algorithm-driven dosing to manage closed-loop glycemic control for people living with type 1 diabetes. Advances in wearable glucose sensing and delivery hardware have supported increasingly compact systems, yet running advanced artificial intelligence dosing algorithms directly on wearable hardware has remained technically challenging. In addition, existing automated delivery architectures frequently depend on external computing hardware such as smartphones or dedicated standalone controllers to execute their core dosing algorithms, which introduces communication dependencies into continuous closed-loop medical operations.

To address these constraints, researchers Jina Lee and Sung Min Park at Pohang University of Science and Technology in the Republic of Korea developed an on-device insulin dosing controller based on the soft actor–critic reinforcement learning algorithm. Their design aims to execute insulin dosing decisions directly on resource-constrained automated insulin delivery hardware. By processing continuous glucose monitoring signals locally on the device itself, the proposed framework eliminates the operational reliance on external communication links or secondary computing hardware during closed-loop management.

Deploying complex reinforcement learning models directly onto embedded hardware presents substantial computational obstacles. To facilitate embedded deployment on a resource-limited platform, the researchers engineered a compressed soft actor–critic architecture equipped with a lightweight encoder. The compression pipeline leveraged quantization-aware training combined with knowledge distillation techniques. Alongside model compression, the system integrated an online adaptation mechanism designed to deliver individualized control tailored to specific patient requirements.

The engineering team implemented the resulting controller directly onto a resource-constrained microcontroller. They evaluated its performance through an experimental hardware-in-the-loop testing framework that utilized virtual patient profiles derived from an established type 1 diabetes simulator. This hardware-in-the-loop configuration allowed the researchers to measure both algorithmic glycemic control performance and physical device execution characteristics under realistic embedded operational constraints.

During in-silico simulation trials, the proposed controller demonstrated effective blood glucose regulation, achieving an average time-in-range of 83 percent across virtual patients. At the same time, the embedded microcontroller sustained continuous real-time on-device execution throughout the testing evaluation. Physical device measurements confirmed that the unit operated with millijoule-level energy consumption per individual control step, demonstrating low computational and power overhead during active closed-loop regulation.

The authors reported that these hardware-in-the-loop and in-silico simulation results establish the technical feasibility of running real-time reinforcement learning-based insulin dosing and personalization on a resource-constrained embedded platform. However, the researchers emphasized that the current findings reflect simulation and embedded hardware bench testing. Comprehensive clinical efficacy and patient safety will require further formal validation before such on-device learning algorithms can be translated to human clinical applications.

The development of automated insulin delivery systems for artificial pancreas applications addresses key architectural constraints associated with external hardware dependencies. Deploying advanced artificial intelligence dosing algorithms directly onto resource-constrained wearable devices presents significant technical challenges, leading many existing systems to rely on external hardware such as smartphones or dedicated controllers to execute dosing routines. This external reliance introduces additional communication dependencies into continuous closed-loop glycemic management. By implementing a soft actor–critic reinforcement learning algorithm directly on an embedded microcontroller, the system processes continuous glucose monitoring inputs and calculates dosing decisions locally. To make this feasible on resource-constrained embedded platforms, the architecture incorporates a lightweight encoder compressed through quantization-aware training and knowledge distillation, enabling local computation without reliance on external computing devices.

Beyond model compression, the embedded system integrates an online adaptation mechanism designed to deliver individualized control for people with type 1 diabetes. In evaluation, the researchers tested the controller within a hardware-in-the-loop framework using virtual patient profiles from a type 1 diabetes simulator. Across these in-silico simulation trials, the on-device controller achieved an average time-in-range of 83 percent while maintaining continuous real-time execution. Furthermore, hardware measurements showed that the microcontroller operated with millijoule-level energy consumption per individual control step. The authors emphasized that while these hardware-in-the-loop and in-silico findings demonstrate the technical feasibility of real-time reinforcement learning and personalization on resource-constrained embedded systems, clinical efficacy and safety require further validation.

Subject of Research: On-device reinforcement learning controller for automated insulin delivery systems

Article Title: On-device reinforcement learning-based insulin dosing controller for fully automated insulin delivery

Article References: Lee, J., & Park, S.-M. (2026). On-device reinforcement learning-based insulin dosing controller for fully automated insulin delivery. Frontiers in Digital Health, 8, 1967489. https://doi.org/10.3389/fdgth.2026.1967489

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1967489

Keywords: automated insulin delivery, artificial pancreas, reinforcement learning, soft actor-critic, type 1 diabetes, continuous glucose monitoring, microcontrollers, embedded systems, on-device AI, knowledge distillation, quantization-aware training, hardware-in-the-loop, glycemic management, in-silico testing, closed-loop control