GTRI

3D Vision and Control of EOD...

3D Vision and Control of EOD Robots using Visual Servoing


Figure 1 Video Demonstration

The function of EOD robots would benefit from a degree of semi-autonomous operation. The ability to control the robot to accomplish a given task is challenging given the difficulty of converting 2D camera data into usable data for the operator. The data is difficult to understand for one major reason: the user lacks a clear understanding of the relationship between the camera data, the real world, and the motions of the robot. Visual servoing is a robot control method that uses camera feedback to determine how to actuate a robot’s joints in order to achieve a desired position and orientation (i.e., pose) with respect to a given target object. The method requires neither precise knowledge of the robot geometry nor any camera calibration.

Work at the Georgia Tech Research Institute (GTRI) has focused on an implementation of technologies based on visual servoing to (1) guide a robot arm to a prescribed pose with respect to a feature of interest (like a window opening) and then (2) provide a more intuitive user interface for finding and manipulating a discovered object. Experiments based on a common scenario of peering inside an unknown opening (a mocked-up car window, in our case) to look for suspicious objects are reported. It is demonstrated that using visual servoing results in a more intuitive and efficient operator experience.


Figure 2 Guide Robot to Window Task


Figure 3 Navigate Within Car


Results to date have been very promising. For sub-task of guiding the robot to the window of a vehicle, the line of sight between the volunteer operators and the robot was removed. They used the video feed from the three fixed cameras for feedback; i.e. it was a teleoperation scenario. The volunteers performed the task of moving the robot end-effector from several different starting positions to the window center. On average, the visual servoing controller was able to move the robot to the target in less than half the time required by the human operators.

A further test was performed for grasping and manipulation of an object within the scene. Early results indicated that while this is a feasible operation to perform for the operator when he has use of the visual servoing algorithm, to process the transformation from eye-in-hand image space to joint coordinates, it is so difficult as to be virtually impossible when he must perform this transformation on his own, i.e. when he issues commands directly in joint space. In this case, the application of uncalibrated visual servoing to joystick operation, based off an eye-in-hand camera, gave abilities to the operator that would otherwise not have been possible.