[Paper Review] Initial Experiments on Learning-Based Randomized Bin-Picking Allowing Finger Contact with Neighboring Objects

Kensuke Harada, Weiwei Wan+4 more|arXiv (Cornell University)|Jul 11, 2016

Robot Manipulation and Learning10 references3 citations

TL;DR

This paper proposes a learning-based randomized bin-picking method that allows finger contact with neighboring objects by training a discriminator on visual point cloud data of surrounding objects to predict grasp success. Using linear SVM and random forest classifiers, the method achieves significantly higher success rates—up to 100% when rejecting predicted failures—by leveraging learned contact configurations to improve robustness over conventional collision-free approaches.

ABSTRACT

This paper proposes a novel method for randomized bin-picking based on learning. When a two-fingered gripper tries to pick an object from the pile, a finger often contacts a neighboring object. Even if a finger contacts a neighboring object, the target object will be successfully picked depending on the configuration of neighboring objects. In our proposed method, we use the visual information on neighboring objects to train the discriminator. Corresponding to a grasping posture of an object, the discriminator predicts whether or not the pick will be successful even if a finger contacts a neighboring object. We examine two learning algorithms, the linear support vector machine (SVM) and the random forest (RF) approaches. By using both methods, we demonstrate that the picking success rate is significantly higher than with conventional methods without learning.

Motivation & Objective

• To improve the success rate of randomized bin-picking in cluttered environments where finger contact with neighboring objects is unavoidable.
• To address the limitation of conventional methods that require collision-free grasping postures, which often fail to find feasible solutions.
• To develop a learning-based discriminator that predicts grasp success based on visual features of neighboring objects and contact configurations.
• To evaluate the effectiveness of different learning algorithms—linear SVM and random forest—in predicting successful picks under contact conditions.
• To enable practical, robust bin-picking in industrial settings by reducing reliance on precise part feeding or manual arrangement.

Proposed method

• A two-fingered robot gripper performs randomized bin-picking on randomly stacked objects, with finger contact to neighboring objects allowed during approach.
• Visual point cloud data of the object pile is captured and used to extract features representing the spatial distribution of neighboring objects.
• A discriminator is trained using either linear SVM or random forest to classify whether a grasp will succeed even if a finger contacts a neighboring object.
• The discriminator uses a high-dimensional feature vector derived from the point cloud distribution to model complex contact configurations.
• Grasping configurations are generated and filtered using inverse kinematics and collision checks, with the discriminator used to pre-screen feasible and likely successful poses.
• The system selects grasping postures from a database of candidate configurations, prioritizing those predicted as successful by the trained discriminator.

Experimental results

Research questions

• RQ1 Can a learning-based discriminator predict successful grasps even when finger contact with neighboring objects occurs during bin-picking?
• RQ2 How do different learning algorithms—linear SVM and random forest—compare in predicting grasp success under contact conditions?
• RQ3 Does using a high-dimensional feature vector improve the accuracy of grasp success prediction in cluttered environments?
• RQ4 To what extent can the proposed method improve the picking success rate compared to conventional collision-free planning?
• RQ5 Can the system achieve higher reliability by rejecting predicted failure cases, even at the cost of reduced attempt frequency?

Key findings

• The random forest-based discriminator achieved a 92.9% success rate when rejecting predicted failure cases, significantly improving upon conventional methods.
• The linear SVM-based discriminator achieved a 100% success rate when only predicted successful cases were attempted, indicating high reliability in filtering failures.
• The random forest method outperformed the linear SVM, identifying 78.0% of trials as successful compared to 66.0% for SVM, suggesting better generalization with high-dimensional features.
• Of 50 trials, 42 succeeded when using the random forest, with 39 of those correctly predicted as successful, demonstrating strong predictive accuracy.
• The system successfully performed picks even when fingers contacted neighboring objects, as confirmed by experimental snapshots and swept volume analysis.
• The method reduced reliance on heuristic rules and demonstrated feasibility for extension to multifingered hands in future work.