Abstract
One of the main challenges in vision-based grasping is selecting feasible grasp regions for novel objects. While Convolutional Neural Networks (CNNs) offer high accuracy, they often require significant computational power and massive labeled datasets. This paper presents an unsupervised learning algorithm that identifies grasp regions using k-means clustering and a novel "Grasp Decide Index" (GDI) for pose selection. Experiments in cluttered environments using Amazon Robotics Challenge objects demonstrate the algorithm's robustness and domain independence, achieving high success rates without external labels or prior object knowledge.
Introduction
Traditional vision-based robotic grasping relies heavily on Deep CNNs, which are notoriously "data-hungry" and often domain-dependent. A model trained on a specific background or set of objects may fail when the environment changes. To address this, the authors propose a computationally lightweight, unsupervised framework that requires no labeled training sets. By focusing on geometric patterns rather than visual classification, the system remains robust across different lighting conditions and backgrounds.
Methodology
The proposed system treats grasping as a geometric pattern-recognition task through a hierarchical algorithm:
1. Sampling and Filtering
The system generates random potential grasp poses across the image plane. Using depth data (Z-values) from an RGB-D sensor, it immediately filters out "no-object" regions (like the floor) and poses likely to cause collisions.
2. K-Means Clustering and Axis Assignment
To localize grasp configurations onto actual object surfaces, the algorithm applies k-means clustering to the filtered points. For each cluster, a "major axis" is calculated to align the gripper with the object's physical structure.
3. Grasp Decide Index (GDI)
The core innovation is the GDI, a metric used to rank grasp candidates. It evaluates the "positive deviation" of sampled pixels within the gripper's footprint relative to the palm's center. This effectively prioritizes poses that minimize the probability of finger collisions with surrounding clutter, essentially turning grasping into a collision-minimization problem.
Experimental Results
The algorithm was evaluated using 50 diverse objects, including items from the Amazon Robotics Challenge (ARC) 2017 and common household goods. Testing across different backgrounds—such as red bins and wooden tables—confirmed its domain independence.
- Cluttered Environments: Achieved an average grasp success rate of 88.18%.
- Isolated Objects: Achieved an average success rate of 95.55%.
The system performed comparably to state-of-the-art supervised models but without the need for the 40,000+ training trials those models typically require.
Conclusion
By prioritizing geometric analysis and collision avoidance over supervised visual classification, this unsupervised approach offers an efficient, lightweight solution for industrial grasping. While the method is sensitive to initial sampling density, its ability to run on a single standard computer while maintaining high accuracy in clutter marks a significant step toward autonomous robotic manipulation in novel environments.
Summary updated