GRASPA 1.0 features several scores in order to fairly evaluate a grasp planner algorithm regardless of the possible limitations of test platform.
The computation of each score requires specific data to be collected on the robot. Hereafter, we summarize the procedure to follow. A sample code for data collection on the iCub is available in GRASPA-test.
- Put your robot in front of the printed layout:
- Estimate the pose of the board reference frame:
- Ask the robot to reach all the desired reachability poses:
- Acquire the poses actually reached by querying forward kinematics.
- Save the received poses using this template and following this procedure.
- Put your robot in the same pose used during the reachability test:
- Estimate the pose of the board reference frame:
- Ask the robot to reach all the desired camera calibration poses:
- Acquire the poses actually reached by using vision:
- Save the received poses using this template and following this procedure.
- Put your robot in the same pose used during the reachability and camera calibration tests:
- Estimate the pose of the board reference frame:
-
Using your grasp planner, compute and make the robot execute 5 grasps for each object in the layout. You can benchmark your algorithm in isolation:
Or in the clutter :
on condition that the same modality is kept for each object and layout.
-
If the robot grasps the object, make the robot execute the trajectory for assessing the grasp stability defined within GRASPA 1.0:
- Store all the data using these templates and following this procedure.
Once all the data for one layout have been acquired, the scores can be computed following these instructions. The procedure needs to be repeated for each layout defined by the benchmark.