back to list

Project: GPU accelerated reactive collision avoidance for robotic control

Description

We want to use moving least squares interpolation for reconstruction of an environment which is used in control of a collaborative robot (Cobot). A cobot is a robot designed to collaborate in a shared space with humans to assist in certain tasks. In order for these robots to avoid collisions with any objects or humans, they need an occupancy map of their environment. This map can then be used in the control loop of the cobot in order to avoid these objects.

A controller for a cobot was designed in a previous project which uses a smooth implicit function as the input for controlling the cobot. Two different previous projects have been working on creating these smooth implicit functions. The smooth implicit functions are generated from a point cloud captured by RGB-D cameras, which capture the environment of the cobot. The limitation of the current implementation of the processing pipeline, needed for the creation of a smooth implicit function, is that it is limited with respect to input resolution. This means that the resolution of the cameras must be fairly low in order to be able to process all points in time.

In this project, we want to use an alternative point interpolation method, needed to create a smooth implicit function, which has the possibility to be implemented on a graphical processing unit. This can improve the performance of the processing pipeline which can enable the use of a higher resolution point cloud. By using a higher resolution point cloud, we can generate a more detailed reconstruction of the environment. During my preparation phase, I have researched multiple methods for improving the performance of the processing pipeline with respect to the current implementation. The main bottleneck of the current processing pipeline seems to be the point interpolation method. Moving Least Squares (MLS) interpolation is a promising alternative to the current interpolation method, which is compactly supported radial basis function (CSRBF). A drawback of MLS interpolation is that it does need normal information at each point in the point cloud, which is not captured by the depth cameras which are used in this project. This is not needed in the current implementation method. This means that these surface normal need to be approximated before MLS interpolation can be used. Multiple methods for estimation of surface normals have been researched, which can be compared on accuracy and processing time.

Details

Student

MT

Mark Tiemersma

Supervisor

Andrei Jalba

Link

Thesis