A containerized simulation platform for robot learning peg-in-hole task

In this paper, we build a simulation platform for robot learning peg-in-hole(PiH) task for to study the strategy of inserting the pegs with different geometry features in PiH task with tele-operation. PiH task as a typical assembly task in the industrial field has been widely researched. Though many researches proposed some general solution for PiH, most of them only rely on accurate force control can be achieved or the environment is structured. In the unstructured environment, it is still a huge challenge. And different sizes and shapes of pegs will significantly increase the difficulty of operation even human-in-loop method because of force and torque introduced from the contact environment and uncertainty from vision, many previous strategies cannot be adapted to these situations. Recently, machine learning method has been achieved many successful applications on robotics which can adapt on different situations with many uncertainties, but making robots learning in the real world still needs more setup, and it also may destroy the robots. Our simulation platform which based on state of art ROS and Gazebo and shipped with Docker and Weave virtual network provides a reproducible and easily deployable platform for robot learning the PiH task. And we also include a tele-operation method for the human operator to tele-operate the simulation robot with force feedback during the peg is approaching to the hole which will enable robot learning trajectory execution from human demonstrations.