Robotic Grasping With Semantic CNNs and Spatial Transformers

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current robotic systems face challenges in accurately detecting and grasping objects using end effectors due to limitations in real-time grasp detection and semantic understanding of objects, especially in diverse environments.

Innovation Solution

The development of deep machine learning methods involving convolutional neural networks (CNNs) and spatial transformer networks (STNs) to predict successful grasp probabilities and spatial transformations, combined with semantic neural networks for object classification, enables robots to improve their grasping abilities by analyzing images and motion data to determine optimal grasping strategies.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep neural networks are used to predict grasp success probability and spatial transformations, then grasping accuracy is improved, but computational complexity and processing time increase

Engineering Contradiction:
Improvegrasping accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the grasp detection task into multiple specialized neural networks: a grasp detection network for predicting success probability, a spatial transformer network for generating transformation parameters, and a semantic network for object classification. Each network processes specific aspects of the grasping problem independently, improving overall accuracy while managing computational complexity through functional decomposition.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary actions by pre-training neural networks on large datasets of grasp attempts and object images before deployment. The networks are pre-computed with learned parameters for predicting grasp success, spatial transformations, and semantic features, enabling fast real-time inference without requiring complex computations during actual grasping operations.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If multiple neural networks are deployed for real-time grasp detection and semantic understanding, then object manipulation accuracy is improved, but system response time deteriorates

Engineering Contradiction:
Improveobject manipulation accuracyVSAvoidsystem response time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system merges multiple neural network functions into an integrated architecture where the grasp detection network, spatial transformer network, and semantic network work together in a unified pipeline. This allows parallel processing of grasp prediction, spatial transformation, and semantic classification, reducing overall response time while maintaining high manipulation accuracy through coordinated multi-network operation.

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If training data from multiple sources is used to train semantic networks, then semantic understanding capability is improved, but data processing complexity increases

Engineering Contradiction:
Improvesemantic understanding capabilityVSAvoiddata processing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The semantic network is designed with universal multi-functionality to process diverse input data from multiple sources including robotic grasp attempts, image databases, and other computer vision datasets. The network architecture handles various data formats and types uniformly, improving semantic understanding capability while managing data processing complexity through a single versatile processing pipeline that adapts to different data sources.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3414710B1Deep machine learning methods and apparatus for robotic grasping
Publication Date: 2022.11.02 GOOGLE LLC
  • EP3414710B1 patent drawingFigure 1A
  • EP3414710B1 patent drawingFigure 1B
  • EP3414710B1 patent drawingFigure 2

AI summary

Deep machine learning methods and apparatus related to manipulation of an object by an end effector of a robot. Some implementations relate to training a semantic grasping model to predict a measure that indicates whether motion data for an end effector of a robot will result in a successful grasp of an object; and to predict an additional measure that indicates whether the object has desired semantic feature(s). Some implementations are directed to utilization of the trained semantic grasping model to servo a grasping end effector of a robot to achieve a successful grasp of an object having desired semantic feature(s).