Aligned Grasping: Wooden Blocks

[KomputerMaster64]

Hello team,

Your work has been helping me in grasping the open-world objects for the pick and place operation.

Recently, I've been trying to use the grasping methods for more precision based grasping. I am using the Franka Emika Panda Robot with a workbench that has wooden cubes (blocks). The pick and place operation involved in this case needs to be more precise since I want to build complex structures with the blocks.

I have extensively tried AnyGrasp and GraspPoseDetection (GPD), however they don't seem to capture the geometry of the objects. I want the gripper to get aligned with the block's edges for better object orientation awareness. AnyGrasp seems to work well with other open set objects like banana, spectacles, bottles etc.

I have also been looking into GraspGPT and GG-CNN lately.

Further, I got some insights from the GraspVLA team (their code is yet to be released):

1. AnyGrasp employs force closure to assess grasp stability, which theoretically means that grasps misaligned with the edges should receive lower stability scores. However, the model's effectiveness can be influenced by the training data and the specific geometries of the objects involved.

2. AnyGrasp relies on sampling methods which may NOT be allowing it to continuously refine the grasp until it reaches a stable configuration to potentially reducing the likelihood of missing optimal grasps due to sparse sampling.

__ AnyGrasp Results

__ GPD Results