Grasping by Hanging: a Learning-Free Grasping Detection Method for Previously Unseen Objects

TL;DR

The paper tackles robust grasping of previously unseen objects without relying on training data. It introduces Grasping-by-Hanging (GbH), a learning-free three-stage framework that identifies hangable structures on an object's mesh, generates 6D grasp poses by aligning a modified parallel gripper with hanging directions, and evaluates candidates with a differentiable score that balances direction toward anti-gravity and free-space safety. Key contributions include a practical hangability-based grasping pipeline, a collision-aware generation and a principled evaluation metric, and real-world validation showing superior performance over a learning-based baseline, especially for thin and flat items. The method holds practical significance for rapid, data-efficient robotic manipulation in environments with diverse, unknown objects, and sets the stage for integration with regular grasping and shape-completion techniques to broaden universality and stability.

Abstract

This paper proposes a novel learning-free three-stage method that predicts grasping poses, enabling robots to pick up and transfer previously unseen objects. Our method first identifies potential structures that can afford the action of hanging by analyzing the hanging mechanics and geometric properties. Then 6D poses are detected for a parallel gripper retrofitted with an extending bar, which when closed forms loops to hook each hangable structure. Finally, an evaluation policy qualities and rank grasp candidates for execution attempts. Compared to the traditional physical model-based and deep learning-based methods, our approach is closer to the human natural action of grasping unknown objects. And it also eliminates the need for a vast amount of training data. To evaluate the effectiveness of the proposed method, we conducted experiments with a real robot. Experimental results indicate that the grasping accuracy and stability are significantly higher than the state-of-the-art learning-based method, especially for thin and flat objects.

Figures (7)

• Figure 1: Examples of a human picking objects with convenient parts for grasping.
• Figure 2: The workflow of the proposed method. The entire procedure includes 3D scanning, hangability detection, grasping detection and grasping evaluation. The hanging position and corresponding hanging direction are marked as red point and black arrow, respectively. Grasping poses are denoted as red lines.
• Figure 3: (a) The image of the gripper used in this work. (b) The simplified model of the gripper. The length of the finger is denoted as $l_f$, the maximum width of the gripper is denoted as $l_w$ and the length of the horizontal rod is $l_h$. We also denote endpoints of fingers and hand as $\mathbf{p}_1$ to $\mathbf{p}_7$. The middle point between the endpoint of the horizontal rod and the finger on the other side is denoted as $\mathbf{p}_M$. (c) An example of the gripper at the grasping pose.
• Figure 4: (a) The workflow of the hangability detection. The rightmost sub-figure shows the final result. The hanging position is represented as $\mathbf{c}_i$ and the corresponding hanging direction is denoted as $\mathbf{v}_i$. $\mathbf{h}_c$ represents the center of mess. Red lines denote the casting rays in the hanging direction detection and blue points denote contact points between rays and the object. (b) The diagram of the hanging position detection. (c) The diagram of the hanging direction detection.
• Figure 5: (a) The image to show the result of hangability detection of two objects. Hanging centers are marked as red points, and hanging directions are denoted as black arrows. Red lines and green lines represent the rays that with or without intersect with the object, respectively. The intersection points between the rays and the mesh are denoted as blue dots. (b) An example of a parallel grasping. In this case, $\overrightarrow{\mathbf{p}_7\mathbf{p}_6}$ is parallel with $\overrightarrow{\mathbf{h}_i^j\mathbf{c}_i}$, $\overrightarrow{\mathbf{p}_1\mathbf{p}_3}$ is parallel with $\mathbf{v}_i$, the $\mathbf{p}_M$ coincides with the $\mathbf{c}_i$. (c) An example of one vertical grasping. In this case, $\overrightarrow{\mathbf{p}_3\mathbf{p}_1}$ is parallel with $\overrightarrow{\mathbf{h}_i^j\mathbf{c}_i}$, $\overrightarrow{\mathbf{p}_4\mathbf{p}_2}$ is parallel with $\mathbf{v}_i$, the $\mathbf{p}_2$ coincides with the $\mathbf{q}_m$. Points caged by the gripper are marked as green in c.1 and c.2 illustrating the local details of determining $\mathbf{q}_m$. (d) Collision-free parallel and vertical grasping poses after the grasping generation are displayed in d.1 and d.2 respectively.