Dex-Net 3.0: Computing Robust Robot Vacuum Suction Grasp Targets in Point Clouds using a New Analytic Model and Deep Learning (1709.06670v2)

Abstract: Vacuum-based end effectors are widely used in industry and are often preferred over parallel-jaw and multifinger grippers due to their ability to lift objects with a single point of contact. Suction grasp planners often target planar surfaces on point clouds near the estimated centroid of an object. In this paper, we propose a compliant suction contact model that computes the quality of the seal between the suction cup and local target surface and a measure of the ability of the suction grasp to resist an external gravity wrench. To characterize grasps, we estimate robustness to perturbations in end-effector and object pose, material properties, and external wrenches. We analyze grasps across 1,500 3D object models to generate Dex-Net 3.0, a dataset of 2.8 million point clouds, suction grasps, and grasp robustness labels. We use Dex-Net 3.0 to train a Grasp Quality Convolutional Neural Network (GQ-CNN) to classify robust suction targets in point clouds containing a single object. We evaluate the resulting system in 350 physical trials on an ABB YuMi fitted with a pneumatic suction gripper. When evaluated on novel objects that we categorize as Basic (prismatic or cylindrical), Typical (more complex geometry), and Adversarial (with few available suction-grasp points) Dex-Net 3.0 achieves success rates of 98$\%$, 82$\%$, and 58$\%$ respectively, improving to 81$\%$ in the latter case when the training set includes only adversarial objects. Code, datasets, and supplemental material can be found at http://berkeleyautomation.github.io/dex-net .

• The paper introduces a compliant suction contact model combined with GQ-CNN to evaluate suction grasp robustness on non-planar surfaces.
• It leverages a 2.8 million-annotated synthetic dataset to train the system, significantly enhancing grasp planning success rates for complex objects.
• Physical experiments validate the approach with success rates up to 98% on basic objects, showcasing its potential for industrial automation.

Analyzing Robust Suction Grasp Planning with Dex-Net 3.0

The paper "Dex-Net 3.0: Computing Robust Vacuum Suction Grasp Targets in Point Clouds using a New Analytic Model and Deep Learning" presents a sophisticated approach to improve suction grasp planning for vacuum-based end effectors, particularly applicable in robotic manipulation contexts like warehouse automation. The research delineates a novel compliant suction contact model combined with a deep learning architecture, GQ-CNN, to evaluate the robustness of suction grasps.

The crux of the problem targeted by the authors is overcoming the limitations of traditional heuristic-based suction grasp plans, which inadequately handle non-planar surfaces often encountered in real-world industrial applications. Vacuum grippers are advantageous due to their simplicity in hardware and ease of maneuvering within cluttered environments. Nevertheless, planning effective grasps for geometrically intricate objects remains a significant challenge.

Key Contributions

1. Compliant Suction Contact Model: The paper introduces a model that quantifies the capacity of a seal to resist external wrenches using a quasi-static spring system. The model is designed to predict the feasibility of a seal formation between the suction cup and the object surface, accommodating variations in material properties that traditional models generally fail to capture.
2. Robust Wrench Resistance: The model evaluates a grasp's ability to resist disturbing forces such as gravity and applied perturbations to both the object and effector poses. This represents a significant shift from previous models that rely on strict assumptions about the rigidity and material homogeneity of objects.
3. Dex-Net 3.0 Dataset: The dataset comprises 2.8 million annotated synthetic point clouds, which provide resources for training the GQ-CNN. This dataset is generated through an analysis of wrench resistance for 375,000 grasps across 1,500 object models, enabling the identification of robust grasp targets.
4. Physical Robot Experiments: The authors validate the system across different object geometry classes. The results demonstrate high success rates for each category—98% for basic objects, 82% for typical geometries, and an improved 81% for adversarial objects, provided the training set is composed of similar complex objects.

Numerical and Experimental Results

The results from this paper substantiate the proposed model's effectiveness, particularly in handling adversarial objects that traditionally pose challenges for simplistic heuristic planning strategies. The experiments reveal that incorporating adversarial object models into the training data significantly enhances the system's ability to generalize across geometrically complex and non-planar surfaces.

Implications and Future Directions

The implications of this research are manifold. Practically, the model could revolutionarily enhance the autonomy and reliability of robotic systems employed in industrial manipulation tasks, particularly those transformed by the Fourth Industrial Revolution. Theoretically, the paper contributes to the body of work seeking to advance the understanding of grasp mechanics through data-driven methodologies synergistically combined with novel analytic models.

Future work could enrich the model's scope by exploring:

• The incorporation of real sensor uncertainties into the dataset to improve robustness under unpredictable environmental conditions.
• Developing composite models that integrate different robotic grippers for more versatile manipulation strategies—combining vacuum-based grip strategies with parallel-jaw techniques.
• Establishing benchmarks for suction grasping scenarios to allow cross-comparison of emerging models and understandings.

Ultimately, the work lays a compelling groundwork for further developments in robust grasping, through enhanced modeling and machine learning insights, dovetailing applied industrial robotics with theoretical inquiry.