Vision-Based Grasping Benchmark

Benchmark pipeline

The general setup of the benchmark is given in the image:

click on the image to enlarge
  1. The real stereo images are input to the user's grasp-generation method. (more)
  2. Using the access tool, the user's vision method processes the images and generates grasping hypotheses. (more)
  3. The grasps are represented as either:
    1. an ordered set of desired contacts, C = {C1, C2, C3},
    2. by choosing one of the hand pre-grasps and the desired hand pose, or
    3. by directly setting the joint angles and hand pose.
    The provided software translates the grasp representation into the format desired by the grasp simulator. (more)
  4. The grasp is executed by the dynamic simulator. (more)
  5. Results with different quality measures are returned to the user. (more)
  6. The results can be displayed using the analysis tool (more)
Note that b) shows the object representation specific to our baseline method.
Gert Kootstra, Mila Popović, Jimmy Alison Jørgensen, Danica Kragic, Norbert Krüger