面向柔软易损食品抓取的大尺寸视触觉夹爪设计与试验

Design and Experiment of Large-scale Visuotactile Gripper for Soft and Fragile Food Grasping

摘要:

柔软、湿滑和易损食品稳定地无损抓取,是中央厨房自动化装填设备面临的关键难题,针对狮子头等典型食品在夹取过程易滑移、易损及接触力难以实时感知等问题,本文设计了一种大尺寸视触觉柔性夹爪。夹爪采用双指包络式结构和四杆平行驱动机构,以厚度为0.03 mm透明PC薄膜作为柔性接触界面,并通过内置摄像头获取标记点位移和接触区域图像信息,实现对柔软易损食品包络夹取。在感知方法上,构建了基于标记点位移场、散度和旋度特征的物理信息神经网络夹取力估计模型,并结合接触区域分割与光流分析实现微滑移检测。在标定试验平台上同步采集视觉图像与外部力信号,建立“视觉位移-接触力”数据集,并开展力估计模型训练与验证。结果表明,PINN模型在法向力估计中RMSE、MAE和平均相对误差分别为0.095、0.091 N和5.40%,均优于非PINN模型。以狮子头、面包、草莓、香蕉块和豆腐块为对象开展夹取试验,结果表明夹爪能在外力扰动条件下保持包络接触,并通过滑移强度变化反映不同食品界面稳定性,多数样品未出现明显破损;所设计的大尺寸视触觉夹爪能兼顾食品抓取的接触顺应性、力感知能力和滑移检测能力。研究结果可为柔软易损食品的自动化装填提供感知与执行一体化方案。

Abstract:

Stable and non-destructive grasping of soft, slippery, and fragile foods remains a key challenge for automated filling equipment in central kitchens. To address slippage, damage, and the difficulty of real-time force perception during the grasping of typical foods such as lion's head meatballs, a large-scale visuotactile flexible gripper was developed. The gripper featured a two-finger enveloping structure and a four-bar parallel driving mechanism, using a 0.03 mm transparent polycarbonate (PC) film as the flexible contact interface, realizing the enveloping and clamping of soft and fragile food. An embedded camera captured marker displacement and contact-region images. A physics-informed neural network (PINN)-based force estimation model was constructed by using marker displacement, divergence, and curl features, while incipient slip detection was achieved through contact-region segmentation and optical flow analysis. A calibration platform was established to synchronously collect visuotactile images and external force signals, forming a visual displacement-contact force dataset for model training and validation. For normal force estimation, the PINN model achieved an RMSE of 0.095 N, an MAE of 0.091 N, and a mean relative error of 5.40%, outperforming the non-PINN model. Grasping experiments on lion's head meatballs, bread, strawberries, banana pieces, and tofu cubes showed that the gripper maintained enveloping contact under external disturbance, and that slip intensity variations reflected the interfacial stability of different foods. Most samples exhibited no obvious damage. The results demonstrated that the proposed gripper can simultaneously provide contact compliance, force perception, and slip detection, offering an integrated sensing-and-actuation solution for the automated handling of soft and fragile foods.