Frequently Questions & Answers
Frequently Questions & Answers
Jump to a question ：
(1) Better model training efficiency than CNNs;
(2) Better offline prediction accuracy than GNN and CNNs;
(3) Better online servoing performance than GNN and CNNs.
(1) Improve the explainability of graph data fed to the GNN;
(2) Provide the valuable feature of depth information for GNN;
(3) Apply the Delaunay method to improve graph edge connection.
Tactile servoing helps to physically interact and manipulate novel objects, even when the target is not visible. It requires changing the region of contact continuously to explore an unknown environment due to the narrow perceptual field. If given the ability to accurately predict the sensor pose with respect to the object's surface, it enables the robot to map the shape and geometry of an object via tactile servoing, which could provide the basis for more complex manipulation tasks.
The hexagonal pins layout of the TacTip sensor is not a standard convex set. Therefore, some virtual nodes need to be constructed to help with convex set generating to identify the position of the outermost boundary.
The biomimetic skin of TacTip senor is soft and flexible which can generate deformation during contact occurs. The shear in the opposite direction of TacTip movement is coupled with the deformation caused by the contact during the servoing. To prevent it from affecting the model prediction of poses, random shear disturbance was added to the data collecting.
Tac-VGNN enjoyed higher-dimensioned features than classic GNN which lacked depth information. This means Tac-VGNN is more capable of feature extraction. Similarly, CNN (ver 1) applied more filters within each convolutional layer to enhance its feature extraction ability. As a benchmark comparison, CNN (ver 2) used a similar structure to the classic GNN for the sake of fairness.
From the comparison of model test performances, CNN (ver 1) had higher estimation accuracy than CNN (ver 2) both on depth pose Y and angle pose Roll. Therefore, only CNN (ver 1) was selected for the offline tactile servoing experiment.
A PI controller was introduced to correct the sensor pose during the tactile servoing. The value of controller gain will affect the servoing step of radial movement and axial rotation. The corner is the junction between the surface and the curve. It requires the transition process of sensor pose adjustment, which causes fluctuations.
(1) Medical robots for precise following of patient's body which helps in the operation of medical equipment such as ultrasound instruments or surgical tools;
(2) Service robots for cleaning variable-shaped furniture, glass and other items used in private homes, offices and public areas;
(3) Assist robots for rehabilitation for the elderly, sportsmen or disabled by doing massage and functional recovery training;
(4) Industrial robots for precise assembly and product quality control via precisely pose manipulation and haptic perception;
(5) Artistic robots for customised painting service on flat, curved and transparent objects.