+ All Categories
Home > Documents > rss-main-09

rss-main-09

Date post: 30-May-2018
Category:
Upload: willow-garage
View: 218 times
Download: 0 times
Share this document with a friend

of 28

Transcript
  • 8/14/2019 rss-main-09

    1/28

    Milestone 2

    Autonomous Door Opening and Plugging inwith a Personal Robot

    Sachin Chitta, Eric Berger, Gary Bradski, Tully Foote, Brian Gerkey,Stuart Glaser, John Hsu, Kurt Konolige, Eitan Marder-Eppstein, Bhaskara Marthi,

    Conor McGann, Wim Meeussen, Patrick Mihelich, Marius Muja, Vijay Pradeep,Radu Rusu, Melonee Wise, Keenan Wyobrek

  • 8/14/2019 rss-main-09

    2/28

    About Us

    Based in Menlo Park, CA

    Focused on open source personal robotics

    ROS (Robot Operating System)

    PR2 (Personal Robot) Look us up

    http://www.willowgarage.com

    http://sourceforge.net/projects/personalrobots

    http://sourceforge.net/projects/ros

    Coming soon

    The PR2 (beta)

    http://www.willowgarage.com/http://sourceforge.net/projects/personalrobotshttp://sourceforge.net/projects/roshttp://sourceforge.net/projects/roshttp://sourceforge.net/projects/personalrobotshttp://www.willowgarage.com/
  • 8/14/2019 rss-main-09

    3/28

    About Us

    We are open source you can take everything we build

    We build robots that can be punished

  • 8/14/2019 rss-main-09

    4/28

    Goal

    Demonstrate the ability to reliably navigate indoors, open doorsand plug into regular outlets

    Robustness and reliability of hardware and software platform

    26.2 mile marathon

    Over multiple days

    10 doors and 10 plugs

    No failures

  • 8/14/2019 rss-main-09

    5/28

    The Robot 2 compliant 7DOF arms

    counterbalanced

    Omnidirectional base, Pan tilt head

    Parallel jaw grippers

    Fingertip sensors

    Sensors

    Laser scanners (2)

    Base, head

    Stereo camera

    Projective stereo

    Regular cameras

    Mounted on head and forearms

    Computing

    4 Core 2 Duo boxes

  • 8/14/2019 rss-main-09

    6/28

    Modular Platform

  • 8/14/2019 rss-main-09

    7/28

    Tasks

    Indoor navigation

    Planning, sensing, localization and control in crowded 3Dindoor environments

    Door opening Deal with doors in different states (open, closed, half-open)

    Plugging in

    Find regular outlets with no special modifications High-level planning

    Robustness to failure

  • 8/14/2019 rss-main-09

    8/28

    Indoor Navigation

    Indoor map

    Generated offline using gmappingand logged data

  • 8/14/2019 rss-main-09

    9/28

    Indoor Navigation

    AMCL (Adaptive Monte

    Carlo Localization)

    Fuses laser scans with

    wheel odometry and IMU

    information

  • 8/14/2019 rss-main-09

    10/28

    Control

    Goals in map frame, control in odometry frame

    map

    t = 1 t = 2

    localization jump resulting obstacle

    odom

    t = 1 t = 2

    localization jump resulting obstacle

  • 8/14/2019 rss-main-09

    11/28

    Navigation

    Dealing with 3D obstacles- voxel grid

    3D is expensive

    Create a 2D array of integers- each bit represents a height value

    Bresenham ray-trace and mark in the grid using bit-wise & and | operations with

    a mask

    x, y, z3, 3, 4

    x, y

    | =

    Adding obstacles

    & =

    Clearing obstacles

    0, 0

    5, 5

    z-mask z-maskselect column

  • 8/14/2019 rss-main-09

    12/28

    Navigation

    Dealing with 3D obstacles- voxel grid

  • 8/14/2019 rss-main-09

    13/28

    Indoor Navigation

    Global Navigation

    Grid based planner

    Local Navigation

    Dynamic window ortrajectory rollout

  • 8/14/2019 rss-main-09

    14/28

    Navigation Milestone

    Reliably traversed 26.2 miles

    32 hours of actual robot navigation time

    Spread over 4 days

    No failures

  • 8/14/2019 rss-main-09

    15/28

    Doors

    ADA compliant

    Smaller search space for width of door, height of handle, shape of thehandle

    Two tasks Detection

    Opening

    Deal with half-open, closed, open, locked doors.

  • 8/14/2019 rss-main-09

    16/28

    Topological Map

    Derived from metric map and door info

    Divide map into door and openregions

    Have a connector between each pair ofregions

    Graph over connectors

    Double-click to add graphics

  • 8/14/2019 rss-main-09

    17/28

    Door detection

    Laser based

    Tilting laser provides 3Dview of the world

    Prior location estimatefrom the map

    Look for vertical planesthat match shape and sizeof doors

    Pick best candidate match

  • 8/14/2019 rss-main-09

    18/28

    Handle Detection

    Laser based + vision based

    Door handle detected from imageusing a boosted cascadedclassifier of haar features

    Handle locations clusteredacross several frames for a morerobust detection

    Stereo information

    Used to eliminate false positives basedon position and size priors

    Used to compute final handle position

  • 8/14/2019 rss-main-09

    19/28

    Door Opening

    Caging grasp on the handle

    Armpushes forwardon the door (nostiff position control)

    Base attracted to goal on other side

    of door + reactive controller forcollision avoidance

    Constraint on shoulder keeps elbowtucked in

  • 8/14/2019 rss-main-09

    20/28

    Door Opening

    Partially open doors

  • 8/14/2019 rss-main-09

    21/28

  • 8/14/2019 rss-main-09

    22/28

    Finding the outlet

    Far outlet detection

    Outlet candidates obtained by segmentingthe disparity image

    Wall pose computed from the base laser

    scan List of outlet candidates is filtered using

    position, size and orientation priors

    Each image patch is rectified to obtain afrontal view of the outlet

    Outlet identity confirmed using a templatematching algorithm

  • 8/14/2019 rss-main-09

    23/28

    Near Outlet detection

    Camera

    5 Megapixel camera on the head with 16mmfocus

    View angle is ~60 degrees, distance is ~1 meter

    Aim for less than 1.5mm error in hole

    location

    Algorithm

    Run a hole detector across the image

    Look for close pairs of holes

    Filter by a single outlet template Group by LePetit's one way descriptor

    Compute detailed homography using holes andknown geometry

  • 8/14/2019 rss-main-09

    24/28

    Near Outlet detection

  • 8/14/2019 rss-main-09

    25/28

    Plug Localization

    Checkerboard target forpose estimation

    At 1m from camera,relative distance accurateto ~0.5mm

    Region of interest

    Centered at gripper position

    Hardware ROI gives

    consistent lighting from auto-exposure

    Large speed boost

  • 8/14/2019 rss-main-09

    26/28

    Plugging in

    Laser based detection of plugon base

    Outlet detection

    Plug localization Visual differencing to reduce

    error between plug and outlet

    Search for outlet

    Plug in

  • 8/14/2019 rss-main-09

    27/28

    TREX - Executive

    Generative Planning

  • 8/14/2019 rss-main-09

    28/28

    Conclusion

    Milestone 2 was a success

    Milestone 3 focuses on you (potential users of the robot)

    Have users successfully develop interesting applications onthe robot

    Future Work

    Doors (opening away from the robot, more handle types,automatically finding doors)

    Plugging in (different types of outlets, better accuracy)

    Navigation (faster speeds)

    Other tasks

    Tabletop manipulation, dynamic obstacles, human-robotinteraction, object recognition, grasping, etc. etc.


Recommended