The Teenager's Problem: Efficient Garment Decluttering as Probabilistic Set Cover

TL;DR

This paper proposes a Probabilistic Set Cover formulation of the teenager's problem, aiming to minimize the number of grasps that clear all garments off the surface, and explores several depth-based methods, which use overhead depth data to find efficient grasps.

Abstract

This paper addresses the "Teenager's Problem": efficiently removing scattered garments from a planar surface into a basket. As grasping and transporting individual garments is highly inefficient, we propose policies to select grasp locations for multiple garments using an overhead camera. Our core approach is segment-based, which uses segmentation on the overhead RGB image of the scene. We propose a Probabilistic Set Cover formulation of the problem, aiming to minimize the number of grasps that clear all garments off the surface. Grasp efficiency is measured by Objects per Transport (OpT), which denotes the average number of objects removed per trip to the laundry basket. Additionally, we explore several depth-based methods, which use overhead depth data to find efficient grasps. Experiments suggest that our segment-based method increases OpT by $50\%$ over a random baseline, whereas combined hybrid methods yield improvements of $33\%$. Finally, a method employing consolidation (with segmentation) is considered, which locally moves the garments on the work surface to increase OpT, when the distance to the basket is much greater than the local motion distances. This yields an improvement of $81\%$ over the baseline.

Figures (5)

• Figure 1: An instance of the Teenager's Problem in the experimental setup; the work surface is white and the basket is beige, a UR5 industrial robot with a Robotiq parallel-jaw gripper is used, with overhead cameras above. The scale automatically records weight data as experiments are run.
• Figure 2: An example of the segment-based grasp point selection algorithm (grasp orientations not shown). From left to right:(1) The original overhead RGB image. (2) The cleaned segmentation ${\mathcal{M}}$. (3) The partitions such that, each partition has the same 'nearby' set of segments; along with $k=5$ grasp candidates sampled from each partition (red dots). (4) The two grasps (red dots) that constitute the grasp plan, found by solving the MILP.
• Figure 3: An example execution using the segment-based method, where the table is cleared in seven grasps in total. Grasp Plan 1 is generated on the initial scene (red dots show planned grasp positions). Only five of these grasps are executed, before Grasp Plan 2 is generated with the remaining two grasps to clean the table.
• Figure 4: An example execution using the consolidation method. The robot performs sequence of pick-and-place actions on the work surface to consolidate garments. It takes only five transport actions to clear all ten garments off the table.
• Figure 5: The test set of 10 garments, representing a variety of different sizes, weights, textures, colors, patterns, flexibility, and garment classes. Some garments have similar colors to present a challenge for segmentation. We also use long garments, such as the scarf, that present a challenge for grasping.