[Paper Review] Primitive Shape Recognition for Object Grasping

Yunzhi Lin, Chao Tang+3 more|arXiv (Cornell University)|Jan 4, 2022

Robot Manipulation and Learning4 citations

TL;DR

This paper proposes a segmentation-based deep learning pipeline, PS-CNN, that decomposes household objects from depth images into primitive shapes (e.g., cylinders, sticks) to enable robust, shape-aware robotic grasping. By leveraging synthetic data and parametrized grasp families per shape, the method achieves 94.2% success in task-free grasping and 93.0% in task-oriented grasping, demonstrating that explicit primitive shape recognition significantly improves grasp reliability and generalization.

ABSTRACT

Shape informs how an object should be grasped, both in terms of where and how. As such, this paper describes a segmentation-based architecture for decomposing objects sensed with a depth camera into multiple primitive shapes, along with a post-processing pipeline for robotic grasping. Segmentation employs a deep network, called PS-CNN, trained on synthetic data with 6 classes of primitive shapes and generated using a simulation engine. Each primitive shape is designed with parametrized grasp families, permitting the pipeline to identify multiple grasp candidates per shape region. The grasps are rank ordered, with the first feasible one chosen for execution. For task-free grasping of individual objects, the method achieves a 94.2% success rate placing it amongst the top performing grasp methods when compared to top-down and SE(3)-based approaches. Additional tests involving variable viewpoints and clutter demonstrate robustness to setup. For task-oriented grasping, PS-CNN achieves a 93.0% success rate. Overall, the outcomes support the hypothesis that explicitly encoding shape primitives within a grasping pipeline should boost grasping performance, including task-free and task-relevant grasp prediction.

Motivation & Objective

• To investigate whether explicitly encoding primitive shapes in a grasping pipeline improves grasping performance.
• To develop a segmentation-based architecture that decomposes objects into multiple primitive shapes using deep learning.
• To enable both task-free and task-oriented grasping by associating shape-specific grasp families with segmented regions.
• To reduce data inefficiency and annotation burden by using synthetic training data for primitive shape detection.
• To improve generalization and robustness to viewpoint variation and clutter in real-world robotic grasping.

Proposed method

• A deep instance segmentation network, PS-CNN, is trained on synthetic depth images of 6 primitive shape classes (e.g., cylinder, sphere, box) generated via a simulation engine.
• The pipeline segments input depth images into distinct primitive shape regions, extracting point clouds per shape class.
• For each segmented shape, a RANSAC-based shape fitting module estimates pose and parameters (e.g., axis, radius, length) to recover the canonical shape model.
• Each primitive shape is associated with a parametrized grasp family that generates dense grasp candidates based on shape geometry.
• Grasp candidates are scored and ranked using a learned scoring function, with the top feasible grasp selected for execution.
• The system integrates classical grasp planning and feasibility checking, enabling real robot execution.

Experimental results

Research questions

• RQ1 Can explicit primitive shape recognition improve robotic grasp success rates compared to implicit shape learning in deep networks?
• RQ2 How well does a segmentation-based pipeline generalize to novel household objects with complex or composite geometries?
• RQ3 To what extent does the method maintain performance under viewpoint variation and cluttered environments?
• RQ4 Can the pipeline support task-oriented grasping by differentiating functionally meaningful object parts (e.g., handle vs. container)?
• RQ5 How does the performance of the method scale with object complexity, especially when objects are not purely one primitive shape?

Key findings

• The PS-CNN pipeline achieves a 94.2% success rate in task-free grasping, placing it among the top-performing methods in the benchmark.
• In cluttered scenes with up to 5 objects, the method maintains a 93.5% grasp success rate and 100% task completion rate.
• The system demonstrates robustness to viewpoint variation, with consistent performance across different camera angles.
• For task-oriented grasping, the method achieves 93.0% success, with a 4% performance drop compared to task-free settings—significantly better than the 19% drop observed in earlier versions.
• PS-CNN v2 outperforms v1 in distinguishing complex object parts, such as the handle and base of a brush, due to improved training data and shape fitting.
• The method shows strong generalization from synthetic training data to real-world objects, even when objects are not pure primitive shapes.