Grasp generation tools #44
Conversation
| class GraspDirection(Enum): | ||
| """ Enumerates grasp direction types. """ | ||
| UNKNOWN = 0 | ||
| X_POS = 1 | ||
| X_NEG = 2 | ||
| Y_POS = 3 | ||
| Y_NEG = 4 | ||
| Z_POS = 5 | ||
| Z_NEG = 6 | ||
|
|
||
| vec_from_direction = { | ||
| GraspDirection.UNKNOWN: None, | ||
| GraspDirection.X_POS: np.array([1.0, 0.0, 0.0]), | ||
| GraspDirection.X_NEG: np.array([-1.0, 0.0, 0.0]), | ||
| GraspDirection.Y_POS: np.array([0.0, 1.0, 0.0]), | ||
| GraspDirection.Y_NEG: np.array([0.0, -1.0, 0.0]), | ||
| GraspDirection.Z_POS: np.array([0.0, 0.0, 1.0]), | ||
| GraspDirection.Z_NEG: np.array([0.0, 0.0, -1.0]), | ||
| } |
There was a problem hiding this comment.
Not sure how this will then morph into supporting grasps at other angles. Maybe the direction should just be a unit vector, but the enum can help look up some common directions?
Actually, the "pose" portion of the grasp should already capture this as the +Z axis should point towards the gripper from the grasp point.
There was a problem hiding this comment.
Fitting a cylinder to the object would make this easier but I could imagine some objects (like a banana) where a cylinder doesn't work well...
There was a problem hiding this comment.
In an ideal world, we'd provide the option of doing cuboid, cylinder, sphere, etc.
While implementation was tough with the cuboid, I mostly picked it because it makes the number of samples for TAMP pretty small, assuming we're just targeting the center of each face as we do in this PR.
I know the MoveIt grasps package, for example, adds more complexity to this by
- Allowing corner grasps
- Sampling offsets from the center
This PR creates a grasp generation pipeline that takes in the gripper / object dimensions, plus the robot pose relative to the object, and generates a set of feasible grasps.
For now, we'll assume all grasps will point towards the center and axis-aligned with a cuboid object.
The goal for this PR is:
In a follow-on PR, this pipeline will be linked to an example showing a next level of TAMP fidelity, where the planner can only pick up objects given feasible grasp locations sampled from this pipeline.
Here is a visualization of the grasp generation. You can run this with:
Begins addressing #45.