A robotic packing system from Worcester Polytechnic Institute treated the object and the gripper holding it as one collision body, then used force sensing and a final push to close the space left for safe placement. On a Franka Emika Panda arm, the pipeline completed 19 of 20 full packing trials across custom irregular shapes and a household-object set.
Dense-packing algorithms often optimize where an object will fit after release. That can produce a mathematically compact pose that the fingers cannot reach without striking the container or another item, while small perception and contact errors accumulate as more objects are added.
Planning for the fingers as well as the object
The optimizer approximates the object, gripper and surrounding space with groups of spheres, allowing it to check quickly whether a proposed placement would overlap something already there. It tests many horizontal positions and rotations in parallel on a GPU. Packing quality depends on the object's shape after release, but collision and approach checks include the fingers that must carry it into place.
Execution then absorbs errors that geometry alone cannot predict. The arm descends until wrist force sensing detects first contact, releases the object, pushes it along eligible horizontal directions to remove residual clearance, and scans the container again before planning the next placement. That scan refreshes the occupied-space map; it does not fit a new position and orientation to every placed object.
Hardware tests separate density from reachability
The paper's space-utilization measure counts only objects fully within the container's footprint and height limit. It divides their combined volume by the total volume of the object set. This is different from the fraction of the container filled, and from whether each placement motion succeeds without a grasp or collision failure.
With seven 3D-printed objects, the full system finished 14 of 15 orderings and succeeded on 104 of 105 individual placements. Packing the largest object first used space more efficiently than either of two random orders, showing that even a reachable planner cannot recover space lost when early objects form a poor base layer.
A separate sequence used eight objects from the YCB manipulation dataset, including a mug, banana, boxes and cans. All five runs completed, all 40 attempted placements succeeded, and mean space utilization was 82.45%, although the study did not randomize the YCB order.
The ablations exposed two different jobs inside the pipeline. Removing the gripper from the planner caused every optimized-sequence trial to fail and cut the average to 1.4 placed objects, while disabling the post-release push preserved 35 of 35 placement successes but reduced utilization to 70.62%.
Against a heightmap-minimization baseline using the same perception, grasps and execution layer, the full planner averaged 95.59% utilization versus 77.66% at the baseline's reported grid resolution and 82.35% with a finer, matched-budget grid. All three configurations completed five of five runs, so this controlled comparison measures packing density rather than a reliability advantage.
Known objects and preselected grasps
The design turns gripper clearance from a final feasibility check into part of placement search, then deliberately recovers that clearance after release. For catalog-based warehouse work, that division could make compact plans executable without padding every object by a large fixed margin.
The current evidence assumes a registered mesh and one preselected grasp for every item, uses only five trials per test configuration, and does not re-estimate each object's pose after placement. The tests did not establish performance on novel objects, alternative grasps, heavier or deformable goods, or longer sequences.
Sources
- Qin et al., "Gripper-Aware Automatic Dense Packing of Irregular Objects", arXiv:2609.22062v1, accepted at ISRR 2026.
- Full paper in HTML, including hardware results, ablations and limitations.
- Representative Franka Emika arm photograph by Ims, Wikimedia Commons, CC BY-SA 4.0. The image shows the robot family, not the reported experiment, and was not modified.