… scene

GraspGenX sees only the object's points, so it proposes grasps from under or
through the table. Nothing checked the gripper at the grasp pose itself: the
planner checks the path to the pre-grasp pose, the last straight move runs with
collision checking off, and the voxel-map ignore box blanks the table under the
target.

GraspGenXModule gets propose_grasps_in_scene(object, scene, threshold). It runs
the model as propose_grasps does, then GraspGenX's own filter_colliding_grasps
on the model's grasp-frame poses, and returns the clear grasps best first with
how many it removed (SceneGrasps). propose_grasps is unchanged.

The gripper shape is a new GraspGenXConfig.gripper_body: boxes in the sweep
volume's frame. The sweep-volume runtime only has a 1 cm placeholder mesh, and
the sweep volume is just the gap between the jaws, so XARM_GRIPPER_BODY gives
the xArm palm (as wide as its flange), knuckles and fingers, measured from the
URDF meshes at drive_joint = 0. Points 7 mm apart on the box faces go to the
filter; the distance batch is sized to stay under 256 MB.

pick_object asks for the check when grasp_scene_filter is on (default) and the
provider offers it (optional SceneGraspGenSpec ref). The scene is the scene
registration's camera cloud of the latest scan with the target masked out,
cut to 0.15 m around the target's points, minus points within 1 cm of the
target's points and within 0.15 m of a tool tip (the wrist camera sees its own
fingers). The table stays. scene_collision_threshold defaults to 1 cm, the
scene cloud's cell size. Each try and the pick result record
scene_filtered_by_source next to rejected_by_source, summed over retries, and
scene_filter_skipped when no scene cloud was available. Heuristic grasps are
not checked.