SIS: Seam-Informed Strategy for T-shirt Unfolding

A common way is to limit the possible garment configurations by first randomly picking up the garment and then grasping the lowest point [4, 13, 14]. A heuristic method proposed by [13] is to unfold a garment by repeatedly grasping the lowest point of the picked up garment.

This strategy first maps the observed garment configuration to a canonical configuration and then obtains the grasping points based on the canonical configuration [15]. The method proposed in [16] estimates the garment pose and the corresponding grasping point by searching a database constructed for each garment type.

Recently, self-supervised learning frameworks [8, 9, 10, 11, 12, 17] have been used to predict action value maps or ranking scores that indicate the unfolding performance of grasping point candidates. [9, 10, 11] also consider the resulting garment orientation from the prioritized action.

The YOLOv3 was originally designed for object detection and has proven to be accurate and computationally efficient. For the SLE, we formulate the extraction of curved$/$straight seams as an oriented line object detection problem since the YOLOv3 cannot be used directly for seam line segment extraction. The proposed formulation allows any object detection network to extract seam line features.

To extract seams using the object detection network, we first approximate the continuous curved seam as a set of straight line segments. Each straight line segment $\textbf{{f}}^{\mathrm{SL}}$ is described by the seam segment category $j\in\{1,2,3,4\}$ and its endpoint positions $(x_{1},y_{1})$ and $(x_{2},y_{2})$ w.r.t the image frame $\Sigma_{\mathrm{img}}$, i.e. $\textbf{{f}}^{\mathrm{SL}}=[j,x_{1},y_{1},x_{2},y_{2}]$. In this paper, the seam segments are categorized into four categories, namely solid ($j=1$), dotted ($j=2$), inward ($j=3$), and neckline ($j=4$), as shown in Fig. 3 (a) and (b), based on the seam types of the T-shirt used in the experiments.

We then convert the straight line segment $\textbf{{f}}^{\mathrm{SL}}$ into a bounding box $\textbf{{f}}^{\mathrm{S}}$ by using the two endpoints as the diagonal vertices of the bounding box. To represent the orientation of the straight line segment $\textbf{{f}}^{\mathrm{SL}}$ in the bounding box $\textbf{{f}}^{\mathrm{S}}$, we divide each seam segment category into four orientation subclasses $i\in\{1,2,3,4\}$, namely downward diagonal ($i=1$), upward diagonal ($i=2$), horizontal ($i=3$), and vertical ($i=4$), as shown in Fig. 3 (a). This quadrupled the number of seam segment categories. Consequently, each oriented seam line segment $\textbf{{f}}^{\mathrm{S}}\in\textbf{F}^{\mathrm{S}}$ is defined by the oriented seam segment category $s_{i,j}$ and the parameters $x,y,\hat{w},\hat{h}$ of the bounding box, i.e. $\textbf{{f}}^{\mathrm{S}}=[s_{i,j},x,y,\hat{w},\hat{h}]$, where $x,y$ denote the coordinate of the center of the bounding box. $\hat{w},\hat{h}$ denote the width and height of the bounding box.

This bounding-box-based categorization of the seam line segment is used as the labeling rule for the SLE shown in Algorithm 1. To implement the SLE, we introduce:

• •

  $\lambda_{thres}$, a threshold for the width and height of the bounding box in pixels: Note that the computation of the loss function during network training is very sensitive to a tiny or thin object whose area is very small. $\lambda_{thres}$ is introduced in Algorithm 1 to guarantee a minimum size of the bounding box.

• •

  Data augmentation of $\textbf{{f}}^{\mathrm{S}}$ with recategorization: The orientation subclass is not invariant if the image is flipped or rotated during data augmentation. Recategorization is carried out using Algorithm 1 when the image is flipped or rotated.

The proposed SLE can extract both curved and straight seams as a set of seam line segments. The predicted seam segments are superimposed onto the original observation $o_{\tau-1}$ to generate the seam map $o^{\mathrm{S}}_{\tau-1}$, as shown in Fig. 3 (c)

As shown in Fig. 3 (b), three types of seam crossing segments, namely shoulder ($c_{1}$), bottom hem ($c_{2}$), and neck point ($c_{3}$) of the T-shirt, are detected by using both SCD1 and SCD2 to increase the recall of the predictions. The inputs of SCD1 and SCD2 are the original image observation $o_{\tau-1}$ and the corresponding seam map $o^{\mathrm{S}}_{\tau-1}$, respectively, as shown in Fig. 2.

Note that the outputs of SCD1 and SCD2 are merged considering the maximum number of each type of crossing segments according to the confidence of the predictions, since the number of each type of seam crossing segments on the T-shirt is limited.

As shown in Fig. 5, the proposed decision matrix $\textbf{{U}}^{\mathrm{D}}$ is constructed as an upper triangular matrix. In this paper, six types of seam segments are considered, including shoulder, bottom hem, neck point, solid, dotted, and neckline (the inward seam line type is treated as the dotted type in this paper for simplicity). The $(k,l)$ element of $\textbf{{U}}^{\mathrm{D}}$, $u^{\mathrm{D}}_{k,l}$, is the performance score of the combination of the $k^{th}$ and the $l^{th}$ seam segment types, and expressed as follows:

where $M_{k,l}$ is the total number of trials of unfolding demonstrations performed by both human and robot. Note that $k,l\in\{1,2,...,6\}$ and $k\geq l$. For each trial, the average normalized coverage $R_{d,k,l}$ is expressed as follows:

where $T_{d,k,l}$ is the number of episode steps in a trial and $\mathrm{ncov}(o_{\tau})$ is the normalized coverage of the T-shirt from an overhead camera observation $o_{\tau}$ at episode step $\tau$ during a trial, similar to [8, 17], and [20]. The normalized coverage is calculated by:

where the coverage $\mathrm{cov}(.)$ is calculated by counting the pixel number of segmented mask output by the Segment Anything Model (SAM) [22]. The maximum coverage $\mathrm{cov}_{max}$ is obtained from an image of a manually flattened T-shirt.

The decision matrix $\textbf{{U}}^{\mathrm{D}}$ is used to select grasping points from the seam segments extracted by SFEM. Based on the score of $u^{\mathrm{D}}_{k,l}$ in the decision matrix and the types of extracted seam segments, the CSST with the highest score is selected. If there are multiple combinations of candidates belonging to each seam segment type of the selected CSST, the most distant candidate pair is selected as the grasping points.

The initial decision matrix is computed using only the human demonstration data by (4). The same equation is then used to update the decision matrix with subsequent robot executions.

Human hands exhibit more dynamic and flexible movements that allow fine-tuning of the resulting garment configuration, a capability that the robot manipulators lack. In other words, there is a human-to-robot gap in garment manipulation. To bridge this gap, we perform real robot trials to update the decision matrix. This iterative method ensures that the grasping strategy integrates the learned human experience and the real robot explorations.

Note that the unfolding performance can be improved by using two decision matrices based on the normalized coverage of the T-shirt, as shown in Section V-C. Using the same initialized decision matrix, we update the two matrices for intermediate and non-intermediate configurations using robot executions. The decision matrix $\textbf{{U}}^{\mathrm{D}}_{\mathrm{int}}$ in Fig. IV-B (c) is for the intermediate configurations with normalized coverage equal to or greater than 0.4, and the decision matrix $\textbf{{U}}^{\mathrm{D}}_{\mathrm{nint}}$ in Fig. IV-B (b) is for the non-intermediate configurations with normalized coverage less than 0.4. The $\textbf{{U}}^{\mathrm{D}}_{\mathrm{nint}}$ is also used for the initial configurations.

Similar to [8], we implement a Grasp&Fling motion primitive to speed up the process of unfolding the garment. After grasping, the T-shirt is stretched by the robot arms until the stretching force reaches a predefined 1.4 N using the wrist F$/$T sensors attached to both manipulators. The robot then generates a flinging motion that mimics the human flinging motion based on fifth-order polynomial interpolation.

To generate initial configurations of the T-shirt that are fair enough for comparison, we randomly select a grasping point from the garment area and release it from a fixed position 0.78 m above the table using one of the robot arms. During the experiments described in Section V-C and V-D, the robot repeats this motion before each trial until the normalized coverage of the initial configuration is less than 0.4.

The dataset used to train the SLE consists of 360 images with manually annotated labels of seam line segments. Using data augmentation by image rotation and flipping, we obtain 2304 images for training and 576 images for validation with annotated labels.

The dataset used to train SCD1 consists of 1481 original images with manually annotated labels of seam crossing segments. The dataset used to train SCD2 consists of the same 1481 images overlaid with the seam lines extracted by the trained SLE.

The original images in all datasets are captured by the Basler camera. The resolution of the images used for training is 1280 $\times$ 1024, obtained by resizing the original images.

The decision matrix $\textbf{{U}}^{\mathrm{D}}_{\mathrm{init}}$ is initialized with human demonstrations. For each human trial, we set the maximum number of episode steps to $T_{d,k,l}=5$, although two to three steps are usually enough to flatten the T-shirt. Once the T-shirt is flattened, we skip the remaining steps and use the same $\mathrm{ncov}(.)$ as in this step for the remaining steps.

To initialize a $u^{\mathrm{D}}_{k,l}$ for each CSST, the demonstrator is asked to randomize the T-shirt until at least one point pair in the CSST appears. The demonstrator then selects the furthest pair of points as grasping points. In the following steps of this trial, the demonstrator selects grasping points based on the demonstrator’s intuition, and the $u^{\mathrm{D}}_{k,l}$ of the CSST is calculated. The matrix $\textbf{{U}}^{\mathrm{D}}_{\mathrm{init}}$ is initialized as shown in Fig. 5 (a). Each $u^{\mathrm{D}}_{k,l}$ is computed using (4) with the number of trials $M_{k,l}=10$.

The DMIM is a simple, yet efficient and easy to implement scheme for selecting grasping points from candidate grasping points extracted by SFEM. As shown in Section V-C, we found that using two decision matrices based on the normalized coverage of the T-shirt significantly improves the unfolding performance. $\textbf{{U}}^{\mathrm{D}}_{\mathrm{int}}$ is used for intermediate configurations when $\mathrm{ncov}(o_{\tau-1})\geq 0.4$ and $\textbf{{U}}^{\mathrm{D}}_{\mathrm{nint}}$ is used for non-intermediate configurations when $\mathrm{ncov}(o_{\tau-1})<0.4$. $\textbf{{U}}^{\mathrm{D}}_{\mathrm{int}}$ is updated from $\textbf{{U}}^{\mathrm{D}}_{\mathrm{init}}$ with (5) when $\mathrm{ncov}(o_{1})<0.4$ and a trial is completed. $\textbf{{U}}^{\mathrm{D}}_{\mathrm{int}}$ is updated from $\textbf{{U}}^{\mathrm{D}}_{\mathrm{init}}$ with (5) when $\mathrm{ncov}(o_{\tau-1})\geq 0.4$ and an episode step is completed.