grasping-short-axis

grasping-short-axis

Deterministic, geometry-locked grasping with CuRobo. The grasp pose is computed directly from the target OBB — the gripper descends along world −Z with its finger-opening axis snapped to the OBB's shorter horizontal axis, so the jaws close across the narrow dimension of the bar. An optional node then slides the grasp outward along the handle, clear of a heavier attached body. The subgraph builds a per-observation collision world (target excluded), plans a CuRobo trajectory, executes it, finalizes the last waypoint, closes the gripper, and outputs the EE pose at grasp.

Install

This skill depends on the curobo and geometry tool bundles:

export CUDA_HOME=/usr/local/cuda
uv sync --extra curobo --extra geometry   # (pip: pip install -e "open-robot-skills[curobo,geometry]" --no-build-isolation)

When to use

• Elongated targets where grasp orientation matters: pan / pot handles, bottles, tools, utensils.
• A subpart (handle) that protrudes from a heavier body — wire base_obb for the outward slide.
• When a sampled/scored grasp pose holds at close but the object slips during the lift (marginal off-axis contact patch).

When NOT to use

• Symmetric objects (boxes, cans) where any yaw works — use grasping-with-planner.
• curobo not deployed.
• You genuinely want multiple sampled candidates for the planner to choose from — use grasping-with-planner.

Recommended subgraph state flow

10 states, in order:

open → compute_grasp → offset_from_base → approach → observe
     → build_world → plan → execute → finalize → close → grasped

(grasped is the success-marker noop from sg.add_exit("grasped"), with an edge to END.)

State details:

1. open — type: tool, tool: "robot.open_gripper", inputs: { settle_steps: 40 }.
2. compute_grasp — type: script, file scripts/<sg>/short_axis_grasp_pose.py (canonical — do NOT re-emit a ```python block; the bundle materializes it). Inputs: target_obb = Ref("in.target_obb"). Optional z_offset (default −0.04: descend the fingertip 4 cm into the OBB top). Returns {grasp_pose: Se3Pose}.
3. offset_from_base — type: script, file scripts/<sg>/offset_grasp_from_base.py (canonical). Inputs: handle_obb = Ref("in.target_obb"), grasp_pose = Ref("compute_grasp.grasp_pose"), and — ONLY when a body perception was authored and base_obb declared as a subgraph input — base_obb = Ref("in.base_obb"). Returns {adjusted_grasp: Se3Pose}. Safe no-op when base_obb is absent.
4. approach — type: script, file scripts/<sg>/approach_above.py (canonical). Inputs: target_position = Ref("offset_from_base.adjusted_grasp.position"), rotation = Ref("offset_from_base.adjusted_grasp.rotation"), target_obb = Ref("in.target_obb").
5. observe — type: tool, tool: "robot.get_observation".
6. build_world — type: script, file scripts/<sg>/build_world.py (canonical). Inputs: observation = Ref("observe"), target_mask = Ref("in.target_mask"), target_obb = Ref("in.target_obb"), target_name = "target".
7. plan — type: script, file scripts/<sg>/plan_grasp.py (canonical). Inputs: world_config = Ref("build_world.config"), observation = Ref("observe"), grasp_poses = Ref("offset_from_base.adjusted_grasp"), target_name = "target". plan_grasp.py auto-wraps the single bare Se3Pose into a one-element list. All four inputs are required.
8. execute — type: tool, tool: "robot.execute_trajectory", inputs: { trajectory: Ref("plan.trajectory"), subsample: 4 }.
9. finalize — type: script, file scripts/<sg>/finalize_trajectory.py (canonical). Inputs: trajectory = Ref("plan.trajectory"). MANDATORY — see the execute → finalize → close hard_rule. Edges: execute → finalize, finalize → close.
10. close — type: tool, tool: "robot.close_gripper", inputs: { settle_steps: 60 }. Edge directly from close to the grasped success marker.

The cross-subgraph output binding:

sg.set_outputs(
    ee_pose_at_grasp=Ref("observe.arms.0.ee_pose"),
    grasp_pose=Ref("offset_from_base.adjusted_grasp"),
)

Wire the exit:

sg.add_edge("close", "grasped")
sg.add_edge("grasped", END)