On the modern assembly line, the posture of flexible products can be distorted or stacked on each other, which is difficult for methods currently in effect to ensure the precision with current methods. To solve this problem, we demonstrate our method on ABB Yumi robot for flexible object detection based on SURF features with color channel prior and marginal feature optimization. Our method achieved 97.99% accuracy, which is 7.67% better than vanilla SURF.

We present a framework that leverages a deep neural network as a feature extractor to learn the low-dimensional representation from high-dimensional input to accelerate the reinforcement learning process. Our experiment shows that the agent vehicle can learn an optimal policy directly from visual inputs achieving significant performance compared with the cutting-edge methods. Moreover, our approach enhances the sample efficiency and policy performance in diverse and sophisticated high-dimensional environments with better robustness.

On the modern assembly line, the posture of flexible products can be distorted or stacked on each other, which is difficult for methods currently in effect to ensure the precision with current methods. To solve this problem, we demonstrate our method on ABB Yumi robot for flexible object detection based on SURF features with color channel prior and marginal feature optimization. Our method achieved 97.99% accuracy, which is 7.67% better than vanilla SURF