Biomimetic grasping is crucial for robots to interact with the environment and perform complex tasks,making it a key focus in robotics and embodied intelligence.However,achieving human-level finger coordination and fo...Biomimetic grasping is crucial for robots to interact with the environment and perform complex tasks,making it a key focus in robotics and embodied intelligence.However,achieving human-level finger coordination and force control remains challenging due to the need for multimodal perception,including visual,kinesthetic,and tactile feedback.Although some recent approaches have demonstrated remarkable performance in grasping diverse objects,they often rely on expensive tactile sensors or are restricted to rigid objects.To address these challenges,we introduce SoftGrasp,a novel multimodal imitation learning approach for adaptive,multi-stage grasping of objects with varying sizes,shapes,and hardness.First,we develop an immersive demonstration platform with force feedback to collect rich,human-like grasping datasets.Inspired by human proprioceptive manipulation,this platform gathers multimodal signals,including visual images,robot finger joint angles,and joint torques,during demonstrations.Next,we utilize a multi-head attention mechanism to align and integrate multimodal features,dynamically allocating attention to ensure comprehensive learning.On this basis,we design a behavior cloning method based on an angle-torque loss function,enabling multimodal imitation learning.Finally,we validate SoftGrasp in extensive experiments across various scenarios,demonstrating its ability to adaptively adjust joint forces and finger angles based on real-time inputs.These capabilities result in a 98%success rate in real-world experiments,achieving dexterous and stable grasping.Source code and demonstration videos are available at https://github.com/nubot-nudt/SoftGrasp.展开更多
基金supported by the Innovation Science Foundation of National University of Defense Technology,China(24-ZZCX-GZZ-11)the National Science Foundation of China(62373201).
文摘Biomimetic grasping is crucial for robots to interact with the environment and perform complex tasks,making it a key focus in robotics and embodied intelligence.However,achieving human-level finger coordination and force control remains challenging due to the need for multimodal perception,including visual,kinesthetic,and tactile feedback.Although some recent approaches have demonstrated remarkable performance in grasping diverse objects,they often rely on expensive tactile sensors or are restricted to rigid objects.To address these challenges,we introduce SoftGrasp,a novel multimodal imitation learning approach for adaptive,multi-stage grasping of objects with varying sizes,shapes,and hardness.First,we develop an immersive demonstration platform with force feedback to collect rich,human-like grasping datasets.Inspired by human proprioceptive manipulation,this platform gathers multimodal signals,including visual images,robot finger joint angles,and joint torques,during demonstrations.Next,we utilize a multi-head attention mechanism to align and integrate multimodal features,dynamically allocating attention to ensure comprehensive learning.On this basis,we design a behavior cloning method based on an angle-torque loss function,enabling multimodal imitation learning.Finally,we validate SoftGrasp in extensive experiments across various scenarios,demonstrating its ability to adaptively adjust joint forces and finger angles based on real-time inputs.These capabilities result in a 98%success rate in real-world experiments,achieving dexterous and stable grasping.Source code and demonstration videos are available at https://github.com/nubot-nudt/SoftGrasp.
