A Multimodal Robust Recognition Method for Grasping Objects With Robot Flexible Grippers

摘要

In light of the critical importance of achieving robust object recognition from multimodal data in robotic operations, this article proposes a precise identification method tailored for the grasping of objects using a multiflexible gripper in scenarios characterized by multimodality, limited samples, and complex environments. Terming the BOSS-MI-ELM algorithm, this approach innovatively extracts [bag-of-SFA-symbols (BOSS)], fusion [association-based fusion (AF)], and classifies [incremental extreme learning machine (I-ELM)] features from multimodal data, facilitating an efficient recognition process. The study employs fiber Bragg grating (FBG) and inertial measurement unit (IMU) as information acquisition components, constructing a multimodal perception system and establishing a corresponding grasping dataset. Through training and testing on this dataset, empirical evidence demonstrates that even with the utilization of only 20% of the dataset, the BOSS-MI-ELM algorithm maintains a classification accuracy of 95.54%. In the presence of Gaussian noise with a mean of 0 and varying standard deviations, as well as different degrees of partial data loss, the proposed method still maintains robust recognition performance. In addition, we have validated the effectiveness of this method in identifying objects grasped at different speeds. Furthermore, comparative experiments were conducted on two publicly available multimodal tactile datasets. The results indicate that the BOSS-MI-ELM algorithm outperforms various baseline models. The extensive experiments collectively demonstrate that this system provides a viable solution for robot object recognition under multimodal tactile perception.