Portal de Periódicos da CAPES

Precision Peg-In-Hole Assembly Based on Multiple Sensations and Cross-Modal Prediction

2022; Springer Science+Business Media; Linguagem: Inglês

10.1007/978-3-031-13841-6_49

1611-3349

Ruikai Liu, Ajian Li, Xiansheng Yang, Yunjiang Lou,

Image Processing Techniques and Applications

Some precision assembly procedures are still manually operated on the industrial line. Precision assembly has the highest requirements in accuracy, which is characterized by a small range of 6D movement, a small tolerance between parts, and is full of rich contacts. It is also difficult to automate because of unintended block of sight, variational illumination, and cumulative motion errors. Therefore, this paper proposes a cross-modal image prediction network for precision assembly to address the above problems. The network predicts the representation vectors of the actual grayscale images of the end-effector. Self-supervised learning method is used to obtain the authentic representation vectors of reference images and actual images during training. Then these vectors will be predicted by combining the reference picture representation, robot force/torque feedback and position/pose of the end-effector. To visualize prediction performance, decoder trained by the above self-supervised network deconvolves the predicted representation vectors to generate predicted images, which can be compared with the original ones. Finally, USB-C insertion experiments are carried out to verify the algorithm performance, with hybrid force/position control being used for flexible assembly. The algorithm achieves a 96% assembly success rate, an average assembly steps of 5, and an average assembly time of about 5.8 s.