Bin-Picking Grasp Planning Using Object Property Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated bin-picking methods face challenges in dealing with unknown objects and require extensive retraining for different gripper setups, leading to inefficiencies and inaccuracies due to the binding of decision-making processes to specific hardware and datasets.
Innovation Solution
A deep neural network system generates object property maps describing measurable physical properties, allowing a separate decision-making algorithm to determine grasp proposals without requiring training data, thus decoupling the decision of where to grasp from the network and enabling flexibility in gripper choice.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a database of objects with pre-determined grasp locations is used, then grasp proposals can be generated quickly for known objects, but the system becomes unable to deal with unknown objects and requires extensive database searching
Solution Approach 1:
The patent extracts the decision-making process from the trained neural network and implements it as a separate analytical algorithm. The neural network generates object property maps (segmentation, curvature, angle, coverage), and the analytical algorithm independently determines grasp proposals based on these maps without being bound by training data decisions. This allows the system to handle unknown objects while maintaining speed.
Solution Approach 2:
The patent segments the grasp proposal generation into distinct functional components: (1) neural network generates property maps, (2) analytical algorithm processes maps to identify graspable regions, (3) grasp proposals are determined based on geometric and physical criteria. This segmentation allows each component to optimize for its specific function without compromising overall system versatility.
2Adaptability or versatility
If analytical methods using mathematical models are used to analyze local features, then the system can handle unknown objects, but the computational cost increases resulting in a trade-off between runtime and accuracy
Solution Approach 1:
The neural network performs preliminary analysis by generating object property maps (segmentation, curvature, angle, coverage) that pre-process the point cloud data. This preliminary action reduces the computational burden on the subsequent analytical algorithm, as it works with pre-computed property maps rather than raw point cloud data, thereby reducing runtime while maintaining accuracy.
Solution Approach 2:
The object property maps serve as intermediaries between the neural network and the analytical grasp proposal algorithm. These maps encode geometric and physical properties in a compact form, allowing the analytical algorithm to efficiently determine grasp proposals without directly processing complex point cloud data, thus reducing computational runtime.
3Speed
If pixel-wise affordance maps are used for suction cup grippers, then grasp proposals can be generated quickly, but the training data binds the network to specific hardware setups requiring retraining for different grippers
Solution Approach 1:
The patent extracts the hardware-specific decision-making from the trained neural network and implements it as a separate analytical algorithm that can be configured for different gripper types. The neural network generates generic object property maps, and the analytical algorithm applies gripper-specific criteria without requiring retraining of the neural network, enabling flexibility across different hardware setups.
Solution Approach 2:
The object property maps generated by the neural network are universal and hardware-agnostic, encoding geometric and physical properties that are relevant for any gripper type. The analytical algorithm can be configured to interpret these maps for different gripper configurations (suction cup, pinch, parallel jaw), making the system universally applicable across multiple hardware setups without retraining.
Data Source
AI summary
Methods and systems for determining a grasp proposal for object picking by a robot gripper are descried wherein the method may comprise capturing an image comprising an object to be grasped by the robot gripper; providing the image to a deep neural network system that is trained to generate an object segmentation map for identifying pixels in the image that are associated with the object and to generate a plurality of object property maps, each object property map linking pixels of the object to information about a predetermined object property; and, determining a grasp proposal for a controller of the robot based on the one or more generated object property maps.