You are here: Foswiki>RoboticsWeb Web>WebTopicList>DARPARoboticsChallenge (06 Mar 2014, dongj2@RPI.EDU)EditAttach

DARPA Robotics Challenge

Introduction
The DARPA Robotics Challenge (DRC) is a robotics competition held by DARPA. The primary technical goal is to develop ground robots capable of executing complex tasks in dangerous, degraded, human-engineered environments. Standing out from the Virtual Robotics Challenge (VRC) and the DRC Trials, our next target is the DRC Finals in December 2014. By then, the Atlas robot will need to attempt a circuit of consecutive physical tasks, with degraded communications between the robots and their operators.

VRC and DRC Trials
The Virtual Robotics Challenge in June 2013 used a simulator called Gazebo to reproduce the dynamic environments a robot may encounter and the simulated Atlas robot was required to accomplish tasks like walking, driving, object manipulation, and so forth. The DARPA Robotics Challenge Trials in December 2013, however, was even more complicated. Those teams who won VRC were offered real Atlas robots, while other teams who did not participate in VRC brought their own. Each robot need to physically compete in eight different tasks: Vehicle, Terrain, Ladder, Debris, Door, Wall, Valve, Hose. With plenty of effort and luck, we were among the top teams for both events in 2013 and moved on to the DRC Finals this December.

Our team and the Atlas robot
Our team TROOPER is composed of Lockheed Martin, University of Pennsylvania and Rensselaer Polytechnic Institute. Lockheed Martin members take charge in the overall software structure of the robot and enable communication between each module in the system. Folks from Upenn work on perception and lower body or whole body control for walking or balancing tasks. Our RPI Robotics Lab, on the other hand, focus on upper body planning for manipulation tasks. We all together form the amazing team.

IntroVideo
  • Introduction of our team
  • OpenDoor
  • Atlas opens and walks through a door
  • Grasp Analysis and Dynamic Test
    Grasp provider is a pre-computed database that contains all the "good grasps" for known objects we will encounter in the competition. "Good" grasps are the ones that the fingers could control all the possible movements of the object, which will allow us to prevent it from sliding away (i.e., the grasp has force closure).

    To build this database, we used a library called GraspIt! It generates a set of random grasps, then uses simulated annealing to search an eight-dimensional space (six for the hand pose and two for, so-called, eigen-grasping actions) to try to find the best grasps for a metric called “grasp energy.” For the purposes of this competition, we use the concept of a target grasp region, which identifies the portion of the object's surface we want to grasp. For instance, if we plan to grasp a drill, we would like the grasp planner to plan the grasps on the handle.

    PickupHose
  • Atlas pickup drill (video link)
  • Graspit Hose
  • Graspit! generates grasps for drill
  • The main weakness of GraspIt! is that grasp selection is based on static analysis. The robot's ability to achieve that grasp in the face of uncertainty and contact dynamics is ignored. Thus our second consideration is to design a grasping action that can achieve a good grasp, and do so reliably, despite uncertainty in the pose of the object. To do this, we choose a grasp suggested by GraspIt! and repeatedly simulate grasping actions with Gazebo. Each time we use the same grasping action, but a different initial pose of the object. The set of poses are chosen to be a discrete approximation of the object's pose uncertainty distribution. Our measure of quality is defined as the percentage of successful grasps from the full set of simulations with pose variations. We plan to develop a similar general-purpose strategy to design robust strategies for performing all of the tasks that Atlas must perform in the VRC.

    Pointing Inverse Kinematics
    We modified traditional iterative inverse kinematic (IK) solvers to support a 'pointing' feature, i.e. let an axis on the end effector frame align with an axis in the base frame of the kinematics chain. This is important because the six-joint Atlas arm has very limited dexterous work space, only about 5% of the entire six-dimensional space. Normal IK solvers require all six degrees of freedom of the end effector to be speficied. Some solvers accept IK solution requests with less than six degrees of freedom specified. Our pointing IK, on the other hand, requires the IK solutions to align the axis while ignoring rotation around that axis. The result is a large gain in the possibility of solving IK for our tasks like aligning hose, turning steering wheel, etc. Moreover, the pointing IK is convenient to use for achieving power grasps, because the open palm can be pushed toward the object until contact, and then the fingers can be closed.
    ArmWorkspace
  • The workspace of the Atlas right arm
  • AtlasDriving
  • Atlas driving behavior
  • Acknowledgement
    This project is funded by Lockheed Martin (LM) and DARPA. We work together with LM and UPenn.

    -- JunDong - 2013-06-05

    -- JunDong - 2014-03-06
    Topic revision: r7 - 06 Mar 2014, dongj2@RPI.EDU
     

    This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
    Ideas, requests, problems regarding Foswiki? Send feedback