Dynamic grasp analysis and profiling of Gazebo

Abstract

In this work, we consider dynamics and uncertainty of the pose of the object during grasp analysis of Atlas robot simulated in Gazebo. Considering dynamics is crucial during grasping process in dynamic environment as well as pose error of the object. The dynamic grasp analysis start with taking grasps from grasp database generated by GraspIt! which is based on static object assumption. We then evaluate the grasps using Monte Carlo simulation. The goal is to identify grasps selected by GraspIt! that are robust to object position uncertainty and accidental bump. We have shown from the result that the success rates of the grasps obtained from GraspIt! differ a lot when in dynamic environment.

Sometimes the simulations are very slow and sometimes are unstable. Thus we find it necessary to determine the bottleneck of the simulator which is another work presented in this thesis. Using a profiling tool, we are able to analyze the program and also the critical section of the code can be identified. We use OProfile as the profiling tool for Gazebo. We have also added our own timing library using the standard counter: rdtsc to ODE and counted the time that each function took in miliseconds within one time step when profiling ODE.