Grasp Execution:

[JemuelStanley47]

Initial tests:

• Correct transform (depth sensor to body frame)
• Get filtered point cloud (.npy -> .npy -> filtered.pcd)
• Store .pcd in data/PCD/
• Run docker container with modified GPD code (no visualization, output 5 grasp poses)
• Get grasp pose, and perform 90 degree about z-axis (Need fixing)
• Execute arm grasp by running script scripts/perform_grasp.py

[JemuelStanley47]

Observations during test grasp execution:

1. Hand-frame orientation = Body Frame orientation : Ideal grasp pose frame is +-90 about gripper-axis

• Rotate gripper-axis rather than performing transformation on grasp pose

2. Gripper/Tool frame is currently at the tip of the gripper. Offset along gripper_z axis would make grasping easier

• Set tool frame offset

3. Gripper padding makes it tough to grasp (less contact area)
4. Robot doesn't exactly reach commanded pose (compared with transforms from snapshot) -> (Separate issue)

• Need a controller : Steady state(PID)
• Explore impedance control Motion Compliance : Impedance control #51

5. Gripper dimensions on GPD

• Modify gripper dimensions on gpd/cfg/eig_config

[JemuelStanley47]

Compute pre-grasp orientation

• Theoretically achieve the target orientation, along a fixed offset along the gripper pose z-axis
• Pre-grasp orientation should also be optimized/ modified(for simplicity) by best-view point selection Best View Point Selection #50

[JemuelStanley47]

Observed that the gripper's 'claw' motion can't exactly be assumed to be 'parallel' as the upper-jaw length > lower-jaw.

[JemuelStanley47]

GPD requires camera_orientation: 0,0,0, so grasp poses won't be filtered according to approach axis if the point cloud is transformed from sensor frame -> body frame.
The modifications required:

1. Get inference from sensor-frame / hand-frame
2. Transform grasp candidates to body frame
3. Execute grasp (might need post-processing for gripper angle orientation)