Table 2.

Types of collision detection techniques and corresponding learning methods used during path planning.

Author(s)	Collision Avoidance Technique	Learning Method	Accuracy
Wang et al. [134]	Zero robot programming for vision-based human–robot interactions, linking two Kinect sensors for retrieval of robot pose in 3D from a robot mesh model.	Wise-ShopFloor framework is used to determine initial and final pose.	N/A
Du et al. [135]	Fast path planning using virtual potential fields, representing obstacles and targets, as well as Kinect sensors.	Human tracking using unscented Kalman filter, for mean and variance determination of a set of sigma points.	Lower avoidance time (>689.41 Hz).
Hongzhong et al. [136]	Preliminary filtering of mesh models to reduce the number of cuboids in experiment. Virtual fixtures known as active constraints used in generating resistive force. Automatic cube tessellation used for 3D point detection and collision avoidance.	Use of oriented bounding boxes (OBBs) and filtering algorithms: Separating Axis Test and Sweep and Prune. Use of field-programmable gate arrays to design a faster GPU system.	Frame rates of 17.5 k OBBs using a bit width of 20, update rate of 25 Hz compared to 1 kHz.
Das et al. [137]	OPML motion planning using standard geometric collision checkers such as proxy collision detectors.	Learning-based Fastron algorithm used to generate robot motion in complex obstacle-prone surroundings.	100-times faster collision detection than C-space modeling.
Torres et al. [138]	Concentric tube robot teleoperation using automatic, collision avoidance roadmaps.	Rapidly exploring random graph (RRG) algorithm aids roadmap construction in maximum reachable insertion workspace.	Tip error between 0.18 mm and 0.21 mm of tip width.
Killian et al. [139]	Multicopter collision avoidance by redirecting a drone onto a planned path; connects random nodes within a search space on a virtual line.	Use of the probabilistic RRT algorithm for collision detection.	Speed of up to 6 m/s.