Purifying Kopi Luwak beans with precise RL-based proximal policy optimization using visual transformer with FRD

Abstract

Among the world’s rarest and costliest coffee beans, luwak beans, after being extracted from the Asian palm civet, a small mammal native to Southeast Asia, and Traditionally beans are harvested, washed, and roasted. Previously cleaning process of luwak beans was a traditional and meticulous practice through hand wash which involves collection, sorting, and pre-washing with water to remove larger pieces of debris and clinging pulp. Cleaning hand-luwak beans with the traditional methods might cause inconsistent quality, potential hygiene concerns, time-consuming and labor intensive. Integrated cleaning units will delicately de-pulp, wash, and dry beans in the proposed method. Features may include color, size, shape, and texture, which are crucial indicators of bean quality. Machine learning algorithms and vision transformers built on picture data will assist the robot’s arms in removing pulp without damaging beans delicately. Controlling drying settings precisely ensures quality and prevents over-drying. The proposed system leverages a Visual Transformer, a powerful image recognition architecture, enhanced with Feature Recombination and Distillation (FRD) for improved accuracy and efficiency. Combining RL with Proximal Policy Optimization (PPO) and a Visual Transformer with Feature Recombination and Distillation (FRD) for visual input processing. Training the RL agent to identify and select high-quality cleaned Kopi Luwak beans based on visual features. They achieved a purification accuracy of 97.57.