GPU-based real-time collision detection for motion execution in mobile manipulation planning

Unified GPU voxel collision detection for mobile manipulation planning

“…The theoretical optimum would be a constant time for collision detection, independent of the number of investigated entities, as this would mean a linear speed up growing with the number of configurations checked in parallel. We showed in [7] that this can almost be achieved with a GPU voxel map.…”

“…This structure is ideally suited for use-case 2 and 3. Further details on implementation and performance can be found in [7], where we showed that maximal parallel access does not achieve the best throughput.…”

“…In [7] we only utilized voxel maps which already offered promising results for scenarios with limited geometrical extent. In this work we wanted to improve the collision detection throughput for larger scenarios.…”

“…At 1 A Kernel is a function that runs with many parallel threads on the GPU and which is launched by the host. 2 [7] for details.…”

“…If a sweep is segmented, also the approximate position of a collision can be determined and used to inform a planner. This approach was extensively investigated in our previous work [7]. Storing swept volumes of motion-sub-plans in voxel lists allows highly efficient motion-primitive planning, as whole trajectories can be evaluated at different positions within the environment map within one collision detection step.…”