[Paper Review] A Configuration-Space Decomposition Scheme for Learning-based Collision Checking

Yiheng Han, Zhao Wang+4 more|arXiv (Cornell University)|Nov 17, 2019

Robotic Path Planning Algorithms20 references4 citations

TL;DR

This paper proposes a configuration-space decomposition scheme that splits the robot's configuration space into DOF-based subspaces, enabling a composite classifier that combines multiple simple classifiers (e.g., SVM, KNN) for improved collision detection accuracy. The method significantly reduces motion planning time—achieving a 30% speedup in a real multi-robot plant phenotyping system—by outperforming single-classifier learning-based approaches in both accuracy and inference speed.

ABSTRACT

Motion planning for robots of high degrees-of-freedom (DOFs) is an important problem in robotics with sampling-based methods in configuration space C as one popular solution. Recently, machine learning methods have been introduced into sampling-based motion planning methods, which train a classifier to distinguish collision free subspace from in-collision subspace in C. In this paper, we propose a novel configuration space decomposition method and show two nice properties resulted from this decomposition. Using these two properties, we build a composite classifier that works compatibly with previous machine learning methods by using them as the elementary classifiers. Experimental results are presented, showing that our composite classifier outperforms state-of-the-art single classifier methods by a large margin. A real application of motion planning in a multi-robot system in plant phenotyping using three UR5 robotic arms is also presented.

Motivation & Objective

• Address the high computational cost of exact collision checking in high-DOF robot motion planning by improving learning-based approximate collision detection.
• Overcome the limitations of single binary classifiers that treat the entire configuration space as a single entity, leading to reduced accuracy in complex environments.
• Develop a composite classifier framework that is compatible with existing learning-based methods and enhances classification performance through spatial decomposition.
• Demonstrate the effectiveness of the method in both synthetic environments with increasing obstacle complexity and in a real-world multi-robot plant phenotyping system.

Proposed method

• Decompose the configuration space $\mathcal{C}$ into multiple subspaces based on the degrees of freedom (DOFs) of the robot, enabling localized classification.
• Construct a composite classifier $F_{\text{comp}}$ by combining multiple elementary classifiers (e.g., SVM, KNN) trained on each decomposed subspace.
• Leverage two key properties of the decomposition: (1) subspaces are more linearly separable, and (2) local classifiers generalize better than global ones.
• Use the composite classifier to perform fast approximate collision checking in sampling-based motion planning (e.g., RRT), reducing reliance on expensive exact collision detectors like FCL.
• Integrate the composite classifier into the RRT framework, where it filters out obviously in-collision or free samples, minimizing the number of exact checks needed.
• Apply a final exact collision check via FCL on ambiguous samples, with local path repair for any in-collision segments.

Experimental results

Research questions

• RQ1 Can decomposing the configuration space based on DOFs improve the accuracy of learning-based collision detection in high-DOF robots?
• RQ2 Does a composite classifier built from multiple elementary classifiers trained on DOF-based subspaces outperform single classifiers in complex environments?
• RQ3 To what extent does the proposed method reduce motion planning time compared to exact collision checking and prior learning-based methods?
• RQ4 Can the composite classifier be effectively integrated into existing sampling-based motion planning pipelines without requiring architectural changes?
• RQ5 How does the method perform in real-world multi-robot systems with complex collision constraints, such as in plant phenotyping?

Key findings

• The composite classifier $F_{SVM}$ achieved a 30% reduction in average motion planning time in a real multi-robot plant phenotyping system, decreasing from 1.0 second to 0.692 seconds per round.
• In synthetic environments with increasing obstacle counts (1 to 8), the composite classifier significantly outperformed single classifiers, achieving a higher success rate within the 3-second time limit.
• The composite classifier reduced the number of required exact collision checks (via FCL) by minimizing false positives and negatives, leading to fewer path repairs.
• The composite classifier had both higher classification accuracy and lower query time than individual classifiers like Fastron, SVM, and KNN.
• The method demonstrated compatibility with existing SBMP frameworks like OMPL’s RRT, enabling plug-in integration without modifying the core planning pipeline.
• The decomposition scheme enabled better generalization, as each elementary classifier operated on a more homogeneous, lower-dimensional subspace, improving overall robustness.