Vision-based grasping focuses on determining successful grasp configurations using input from vision sensors.
Typically, algorithms take RGB and depth images from modern RGB-D cameras and output where and how to grasp a target object in 3D space.
With...
Vision-based grasping focuses on determining successful grasp configurations using input from vision sensors.
Typically, algorithms take RGB and depth images from modern RGB-D cameras and output where and how to grasp a target object in 3D space.
With advances in grasping algorithms and vision sensors, vision-based grasping has become more reliable across a wide range of scenarios.
Grasping models, enhanced by sophisticated engineering, can incorporate physically grounded biases—such as smoothness and center of gravity—to generate high-quality grasps based on object geometry.
At the same time, vision sensors leveraging technologies like stereo vision, LiDAR, and infrared have become more capable and affordable, allowing accurate geometry capture for most objects.
Moreover, the increasing availability of real-world datasets has significantly boosted performance in practical robotic applications.
However, scenes containing transparency and clutter often fail to be correctly recognized by vision sensors, greatly reducing the reliability of vision-based grasping algorithms.
First, transparent objects cause complex sensing failures in depth cameras due to its physical properties.
Second, clutter, where multiple objects are in contact, results in occlusions and unobservable surfaces.
These elements obstruct the acquisition of accurate geometry in a vision-based manner, leading to inaccurate grasps.
In this thesis, I propose a reliable approach for acquiring scene geometry in environments with transparency and clutter.
By leveraging information available from general pretrained vision modules, the method focuses on the aspect of generalization to various scenes.
It extracts and utilizes mid-level representations such as masks and surface normals in spatially structured ways to achieve stable geometric reconstruction.
First, I introduce a data-driven method for reliably obtaining instance masks in cluttered scenes containing transparent objects, along with a corresponding grasping algorithm.
Second, I propose a novel 3D representation based on surface normals to effectively capture the geometry of transparent objects.
Finally, I present an interactive perception pipeline that actively acquires instance-level 3D geometry in cluttered scenes with occlusions.
Furthermore, experiments conducted on a real-world robotic platform demonstrate the potential for practical deployment in real-world scenarios.
To enable robots to effectively replace human labor across diverse scenarios, consistent performance in variable conditions must be ensured.
By addressing transparency and clutter—two major challenges for vision-based grasping—and proposing solutions built on general-purpose vision modules, this thesis aims to improve the stability and robustness of robotic manipulation.