Week 9: Mapping the Limits of Reach — Suction Grasping Under Spatial Variation

Integrating a pneumatic suction paradigm, then stress-testing the pipeline across randomized objects, a 14-cell target grid, and three robot-base poses — and what the failures taught us about workcell layout optimization.

Posted Nov 5, 2026 Updated Nov 5, 2026

By Ahmed Algharib

7 min read

Week 9 — Mapping the Limits of Reach :Pneumatic Suction, Spatial-Variation Testing, and Expert Algorithmic Foundations

Internship: JdeRobot 2026 — Ahmed Algharib Package: vibe_version · Platform: UR5e MuJoCo twin · ROS 2 Jazzy · MoveIt 2 Scope of this week: integrate the pneumatic suction paradigm, answer the mentor’s challenge (vary object, target, and robot-base positions), and ground every architectural choice in the expert literature.

TL;DR

Suction paradigm integrated: surface-normal-aligned approach + proximity-gated weld reuse; decomposed pick sequence (pre-grasp → linear descent → vacuum → linear lift).
Mentor’s challenge executed: systematic ablations over object poses, a 14-cell target grid, and three robot-base configurations.
Key finding: a single fixed base cannot cover the 14-cell grid; two complementary base poses cover all 14 cells → this is workcell layout optimization, not an algorithmic failure.
Literature review completed: classical vs learning vs hybrid approaches, mapped to our exact failure modes, with a study-milestone plan.

1. Pneumatic Suction Gripper Integration

A suction cup, unlike a parallel jaw, tolerates no tilted approach: the tool Z-axis must be collinear with the object’s surface normal (top-down for boxes; apex-targeting for the ball and cylinder).

Decomposed suction sequence (state machine):

Pre-grasp 10 cm above the object along its surface normal.
Linear Cartesian descent until the cup face contacts the surface.
Vacuum engage → proximity-gated weld (reuses the Week 8 GraspAttachmentPlugin: ≤ 3 cm gate, body-based weld, offset captured at activation).
Linear Cartesian lift before any free-space motion.
Transfer → release on the physics step (real gravity).

This keeps the strict orientation constraint inside the planner (Cartesian paths with locked approach vector) instead of hoping joint-space motion preserves the seal.

2. The Mentor’s Challenge: Systematic Spatial Variation

2.1 Test design

Three variation dimensions, one at a time (ablation):

Object positions: randomized pickup poses in a bounded table region (seeded).
Target positions: a 14-cell grid spanning the extended table, including outer-right edge cells at X ≈ 0.967 m.
Robot base positions/orientation: original base (180° yaw), original base (0° yaw), shifted base (0.186, 0.122) m.

Fig 1: The variation-testing workcell. MuJoCo (left): target bins aligned along the table edge forming the 14-cell grid, objects randomized on the tabletop; RViz (right): the identical scene in the MoveIt planning environment.

2.2 Results

2.3 Finding: workcell layout optimization

The UR5e’s nominal reach is 850 mm; the failing cells sit at the kinematic boundary. The two configurations are complementary: their union covers all 14 distinct cells, proving the software stack is capable everywhere and the limitation is purely geometric. A single static base pose cannot cover this grid without either dropping edge targets or violating base clearance — the classic fixed-base workcell trade-off, and the justification for treating base placement as an optimization variable (see §4.5).

3. Safety and Validation Integrity

All edge-cell failures were pre-execution validation rejections (IK/planner/validity checks at ~0.967 m). The robot never attempted an unreachable trajectory: the pipeline filters impossible poses before MuJoCo execution, guaranteeing zero unstable physical behavior during unattended runs.

4. Literature Review: Expert Algorithmic Approaches

4.1 Motion planning — sampling vs trajectory optimization

Ours: OMPL RRT-Connect. Probabilistically complete; benchmark evidence (MotionBenchMaker-style studies) shows RRT-Connect degrades gracefully under pose variation (e.g., 85%→76% success under shelf rotation) with consistent planning times. Alternative: CHOMP/STOMP/TrajOpt yield smoother industrial paths and better maintain normal alignment during approach, but are local optimizers that trap on poor seeds.

4.2 Inverse kinematics — analytical vs numerical

Ours: multi-seed analytical/hybrid (TRAC-IK style). Suction demands near-zero orientation tolerance; multiple seeds reliably converge near workspace boundaries. Alternative: pure numerical (KDL Jacobian pseudoinverse) oscillates or times out at boundaries and wrist singularities (J5 ≈ 0°) — unacceptable at the 0.967 m edge cells.

4.3 Coordinate paradigms — relative vs absolute

Ours: object-centric relative poses. The grasp frame is computed from the live object pose, so the suction approach vector always aligns with the local surface normal under any randomization. Alternative: absolute world-frame hardcoding breaks the moment the cell shifts or objects move.

4.4 Grasp execution — proximity weld vs force-torque feedback

Ours: state-machine proximity weld in MuJoCo (distance gate + normal alignment + eq_active flip inside pre_step). Efficient, deterministic vacuum surrogate. Alternative: FTS pressure-drop feedback is the physical truth on hardware but requires fluid-level modeling to simulate faithfully.

4.5 Advanced strategies for positional variability (from the deep-research survey)

Reachability inversion: precompute the arm’s reachable workspace and invert it to sample base poses (SE(2)) from which a grasp is achievable — directly applicable to our base-pose search.
Policy mobilization (Mobi-π): given a policy trained at a fixed pose, score candidate robot poses in a new scene and optimize the start pose — a data-efficient alternative to retraining.
Augmented Kinematic Representation (AKR / SMMP): treat base + arm + held object as one kinematic chain and plan jointly in the augmented space; reported 90–95% success on long-horizon tasks where decoupled planning fails.
Hybrid TAMP: symbolic task sequencing + sampling-based motion planning + learned skills (e.g., RL decluttering) for cluttered scenes.

Comparison matrix

5. Study Milestones (understanding each expert approach)

M1 — Planners: how sampling builds C-space graphs; why RRT-Connect is probabilistically complete; where TrajOpt’s local minima come from.
M2 — IK: closed-form vs Jacobian pseudoinverse; wrist-singularity geometry of the UR5e; why seeds matter under orientation locks.
M3 — Frames: object-centric vs world-frame math (T_grasp^obj); effect on suction normal alignment.
M4 — Grasp physics: weld constraints vs contact/friction vs FTS; what each can and cannot prove.
M5 — Synthesis: map each approach to our measured failure modes (edge-cell IK rejections, clearance exclusions) and decide what to adopt next.

6. Current Status

✅ Suction sequence integrated and exercised across all variation runs.
✅ 14-cell grid tested under three base configurations; coverage JSON saved per run.
✅ All 14 cells covered by the union of configurations.
✅ Pre-execution safety filter verified (zero physical attempts at unreachable poses).
✅ Literature review + comparison matrix drafted (this document).
⏳ Automated base-pose search and placement-error quantification (next).

7. Next Steps

Base-pose micro-grid search between (0,0) and (0.186, 0.122) m (informed by reachability-inversion ideas) for a single-sweep sweet spot.
Quantify placement error (ε_p, ε_θ) across validated cells.
Large-scale dataset collection with combined pickup+target randomization (manifest CSV).
Continue SOTA review (BC, ACT, Diffusion Policy, Open X-Embodiment) → Week 10 blog.

8. Known Limitations

Suction is idealized (proximity weld): no slip, seal-loss, or fluid dynamics.
Single static base cannot cover the full 14-cell grid (geometric, documented).
Base-pose selection is still manual ablation, not optimized.

Appendix A — Figures

| File | Section | Content | |—|—|—| | images/week9_workcell_14cell.png | 2.1 | 14-cell workcell in MuJoCo + RViz | | images/week9_coverage_terminal.png (optional) | 2.2 | Coverage summary per configuration |

Appendix B — Reproduce

```bash ros2 launch ur5e_mujoco_moveit mujoco_moveit.launch.py ros2 run vibe_version pick_place_relative –object yellow_box –episodes 10 –seed 27

Internship

This post is licensed under CC BY 4.0 by the author.