Smart Grippers for Predictive Object Recognition in Warehousing Robots

Abstract

Smart grippers that integrate predictive object recognition capabilities have emerged as pivotal components in the next generation of autonomous warehousing robots. By anticipating both the identity and grasp affordances of items prior to contact, these systems can significantly reduce mispicks, optimize cycle times, and enhance overall throughput. This paper presents an end-to-end framework combining a soft, sensorized gripper design with a parallel deep learning architecture that fuses visual and tactile modalities. The hardware includes an Ecoflex-based soft gripper embedded with triboelectric nanogenerator (TENG) sensors—both pressure (P‑TENG) and bend (B‑TENG)—that provide rich tactile feedback. On the algorithmic side, we deploy a dual-network pipeline: a YOLOv5 model for rapid object detection and classification, and a GG‑CNN for pixel-wise grasp quality estimation, both trained and fine-tuned on a custom warehouse dataset. A lightweight convolutional network processes the high‑frequency tactile signals to infer object geometric categories. These streams converge in a Bayesian fusion module that dynamically selects optimal grasp parameters (finger spread, approach angle, and force threshold) before actuation. In a 1,000‑trial evaluation across ten item classes in a mock warehouse environment, our system achieved a predictive recognition accuracy of 94.3% and a grasp success rate of 92.1%, reducing mispicks by 37% and improving average pick cycle time by 18% compared to vision‑only baselines. Under challenging conditions—low lighting (< 100 lux) and partial occlusions—the fused approach maintained an 89.7% success rate, whereas vision‑only performance dropped to 78.4%.