Jeannette Bohg
Postal address CVAP/ NADA/ KTH
S-100 44 Stockholm
Sweden
Office addressRoom 609
Teknikringen 14
Stockholm
Phone+46 (0)8 790 7224
e-mail
 
CV
Research
Publications
Videos
Code
I am a Ph.D. student at the Computer Vision and Active Perception Lab at the Royal Institute of Technology. My supervisor is Danica Kragic. I am involved in the EU-funded project Grasp.

My Niklas number is 1.
Research
Top
  • Robotic Grasping
  • Simultaneous Localisation and Mapping
  • Computer Vision
  • Machine Learning
  • Speech Processing
Publications
Top
Conference Papers
Journal Papers
Monographies
Workshop Papers and Invited Talks
Videos
Top
Vision-based Grasp Inference
These videos are demonstrating the vision-based grasp inference system based on the work presented in [ bib].
Here are some of the successful grasps: Here are some of the not successful grasps, i.e., the hand does not manage to pick up the object or picks it up at the wrong point. Integration with closed loop control would help to overcome imprecise calibration and to refine the hand pre-shape.
Multi-Modal Scene Exploration
This video demonstrates the different strategies for multi-modal scene exploration presented in [bib]. Vision and touch sensors were used for incrementally building up a scene representation. Unobserved places were predicted with a Gaussian Process. Through this prediction haptic exploration was guided.
Active Vision System for Building up a Scene Model
In this video, we show the application of our active vision system to the problem of cleaning a table. In this example we consider only known objects. The details are described in [bib]. We show the numerous computational processes necessary to build up a scene representation in which this task can be performed.
Additionally, we also show our offline calibration system used for stereo calibration, hand-eye calibration and the calibration of kinematic chain of the Armar III Robotic head.
Scene Understanding through Human-Robot Dialog
In this video we show a prototypical dialog between a human and a robot to refine the scene model the robot has initially acquired.
Grasp and Motion Planning given Incomplete Observation
Classical grasp and motion planners have been commonly developed in simulation with complete models of the environment and the objects given. A robot however is observing its surroundings through real-world sensors producing incomplete and noisy environment and object models. In this video, we demonstrate an approach on two different robotic platforms in which we bridge the gap between real-world sensing and classical grasp and motion planners. It is based on the work presented in [bib].
Code
Top
ROS Packages
During a two-month internship at Willow Garage, I was involved in the creation of several ROS packages.
  • The package [active_realtime_segmentation] implements the method that was proposed by Mårten Björkman in [bib]. Example code for using this code stand alone and as a ros package is contained.
  • The package [object_segmentation_gui] implements an rviz plugin for interactive segmentation using the real-time segmentation method from above.
  • The [rgbd_assembler] is a helper package porting RGB information from the wide-field cameras of the PR2 to its monochrome narrow-field cameras through 3D reconstruction.
  • The package [fast_plane_detection] extract the dominant plane in a scene in real-time. It is a submodule from the above mentioned segmentation module.

Created by: Jeannette Bohg, Photo by: Sebastian Sussmann Last modified: 04-06-2011 01:55