A new differentiable architecture search method for optimizing convolutional neural networks in the digital twin of intelligent robotic grasping

Convolutional neural networks (CNNs) have been widely used for object recognition and grasping posture planning in intelligent robotic grasping (IRG). Compared with the traditional usage of CNNs in image recognition, IRGs require high recognition accuracy and computational efficiency. However, the existing methodologies for CNN architecture design often rely on human experience and numerous trial-and-error attempts, which make it a very challenging task to obtain an optimal CNN for IRGs. To tackle this challenge, this paper develops a new differentiable architecture search (DARTS) method considering the floating-point operations (FLOPs) of CNNs, named the DARTS-F method, which converts the discrete CNN architecture search to a gradient-based continuous optimization problem and considers both the prediction accuracy and the computational cost of the CNN during the optimization. To efficiently identify the optimal neural network, this paper adopts a bilevel optimization, which first trains the neural network weights in the inner level and then optimizes the neural network architecture by fine-tuning the operational variables in the outer level. In addition, a new digital twin (DT) of IRG is developed considering the physics of realistic robotic grasping in the DT’s virtual space, which could not only improve the IRG accuracy but also avoid the expensive training time. In the experiments, the proposed DARTS-F method could generate an optimized CNN with higher prediction accuracy and lower FLOPs than those obtained by the original DARTS method. The DT framework improves the accuracy of real robotic grasping from 61 to 71%.