Real Time Collision Avoidance with GPU-Computed Distance Maps

TL;DR

The paper introduces a unified, real-time collision avoidance framework for redundant robotic manipulators that leverages GPU-accelerated 3D Euclidean Distance Transforms (EDTs) on a voxelized environment and a compact, offline voxelized robot model built from oriented bounding boxes and minimal spheres. By computing exact EDTs with the Parallel Banding Algorithm (PBA) on the GPU and integrating obstacle and self-collision tasks within a task-priority controller, the approach achieves fast, smooth reactive behavior even in the presence of dynamic obstacles, demonstrated on the Tiago robot with live sensor data. The architecture minimizes CPU-GPU data transfer, uses a Bayesian voxel map for obstacles, and applies a task-oriented regularization to mitigate practical discontinuities during task activation. Experimental results show EDT updates at hundreds of Hz and successful collision avoidance while maintaining end-effector tracking, outperforming CPU-based distance-field methods in speed and scalability for larger workspace representations.

Abstract

This paper presents reactive obstacle and self-collision avoidance of redundant robotic manipulators within real time kinematic feedback control using GPU-computed distance transform. The proposed framework utilizes discretized representation of the robot and the environment to calculate 3D Euclidean distance transform for task-priority based kinematic control. The environment scene is represented using a 3D GPU-voxel map created and updated from a live pointcloud data while the robotic link model is converted into a voxels offline and inserted into the voxel map according to the joint state of the robot to form the self-obstacle map. The proposed approach is evaluated using the Tiago robot, showing that all obstacle and self collision avoidance constraints are respected within one framework even with fast moving obstacles while the robot performs end-effector pose tracking in real time. A comparison of related works that depend on GPU and CPU computed distance fields is also presented to highlight the time performance as well as accuracy of the GPU distance field.

Figures (13)

• Figure 1: Representation of the Tiago robot in the scene: (a) links are offline voxelized and inserted according to the joint state of the robot; (b) the arm of the robot is approximated with spherical volumes, the body highlighted in yellow is used as an obstacle for self-collision avoidance.
• Figure 2: The Proposed Collision Avoidance control scheme. The main components of the proposed system are embedded in the blocks marked with blue. The GPU-Voxels data structure updates and stores occupancy information of the environment from live 3d point cloud sources(\ref{['subsec:obstacle_rep']}). For self Collision avoidance, the robot's link is converted to voxels and stored in a separate GPU-voxels map. Based on the occupancy information of the environment and the robot, the 3D-Euclidean distance is computed at very high rate using the PBA algorithm\ref{['exact EDT']}. This is utilized during the Obstacle and Self-collision Avoidance of the Task-Priority Algorithm described in Section \ref{['obstacle avoidance']}. The ROS control Hardware Abstraction in the final block handles out of the box interface with our Task Priority Control in real time (RT).
• Figure 3: Raw point cloud processed to remove noise and robot part from the scene: Raw point cloud visualization in Rviz containing points coming from the robot gripper (left) and GPU voxels representation (right).
• Figure 4: (a) OBB defined over each link; (b) OBB-Aligned minimal spherical bounding volume Bounding: computed using OBB of each link.
• Figure 5: Pipeline for Self collision minimum distance computation: The voxel model of Links that are not in the ACM are transformed and inserted inside a GPU_voxels map followed by EDT.