Intelligent robotic manufacturing systems are revolutionizing the production industry.These next-generation systems employ robots as actuators,multi-source sensors for perception,and artificial intelligence for decisi...Intelligent robotic manufacturing systems are revolutionizing the production industry.These next-generation systems employ robots as actuators,multi-source sensors for perception,and artificial intelligence for decision-making,aiming to execute routine manufacturing tasks with greater autonomy and flexibility.In footwear manufacturing,sole deburring presents a specific challenge in detecting defects and elaborating deburring paths,which skilled workers traditionally handle.The present research goes beyond solving such problems traditionally with computer vision and hard robot programming.Instead,it focuses on developing a learning structure mimicking human motion planning capability from vision inputs.Like humans who mentally visualize and predict a path before refining it in real-time,we want to give the robot the ability to predetermine the trajectory needed for a finishing task,exploiting only vision data.The system is designed to learn how to identify defects and directly correlate this information with motions by utilizing a latent space representation,transitioning from simple programmed responses to more adaptive and intelligent behaviors.We call it a self-supervised vision-proprioception model,an AI framework that autonomously learns to correlate visual observations to proprioceptive data(end effector trajectories)for effective task execution.This is achieved by integrating a vision-based latent space learning phase(learn to see),followed by a reinforcement learning stage,where the agent learns to associate the latent space with deburring actions in a simulated environment(learn to act).Recognizing the common performance degradation when transferring learned policies to real robots,this research also employs Sim-to-Real methods to bridge the reality gap(learn to transfer).Experimental results validate the whole approach.展开更多
基金partly carried out within the MICS(Made in Italy—Circular and Sustainable)Extended Partnership and received funding from Next-Generation EU(Nos.Italian PNRR—M4 C2,Invest 1.3—D.D.1551.11-10-2022,PE00000004),CUP MICS D43C22003120001.
文摘Intelligent robotic manufacturing systems are revolutionizing the production industry.These next-generation systems employ robots as actuators,multi-source sensors for perception,and artificial intelligence for decision-making,aiming to execute routine manufacturing tasks with greater autonomy and flexibility.In footwear manufacturing,sole deburring presents a specific challenge in detecting defects and elaborating deburring paths,which skilled workers traditionally handle.The present research goes beyond solving such problems traditionally with computer vision and hard robot programming.Instead,it focuses on developing a learning structure mimicking human motion planning capability from vision inputs.Like humans who mentally visualize and predict a path before refining it in real-time,we want to give the robot the ability to predetermine the trajectory needed for a finishing task,exploiting only vision data.The system is designed to learn how to identify defects and directly correlate this information with motions by utilizing a latent space representation,transitioning from simple programmed responses to more adaptive and intelligent behaviors.We call it a self-supervised vision-proprioception model,an AI framework that autonomously learns to correlate visual observations to proprioceptive data(end effector trajectories)for effective task execution.This is achieved by integrating a vision-based latent space learning phase(learn to see),followed by a reinforcement learning stage,where the agent learns to associate the latent space with deburring actions in a simulated environment(learn to act).Recognizing the common performance degradation when transferring learned policies to real robots,this research also employs Sim-to-Real methods to bridge the reality gap(learn to transfer).Experimental results validate the whole approach.
