Robot Grasp Pose Detection with Lightweight AI and 4PCS

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing robot pose estimation in industrial settings is inaccurate and slow, and large AI models are resource-intensive, making them unsuitable for industrial edge devices, especially in scenarios with minimal point cloud overlap or small targets.

Innovation Solution

A method utilizing a lightweight AI model based on a self-attention mechanism with a sliding window for target detection, combined with 4PCS algorithms, to accurately and efficiently determine robot pose by transforming point clouds and calculating transformation matrices.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If transformer-based AI models are used for target detection, then detection accuracy is improved, but device complexity and resource consumption increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidmodel size
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the point cloud data into multiple local regions and processes them separately using a lightweight network, rather than using a large transformer model on the entire point cloud. This divides the complex detection task into manageable local segments that can be processed efficiently on edge devices.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts key local features from point cloud regions and uses only the essential information for detection, rather than processing the complete point cloud data through a large model. This extraction approach reduces computational requirements while maintaining detection accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

2Measurement precision

If large AI models are deployed on industrial edge devices, then detection capability is improved, but speed of reasoning deteriorates

Engineering Contradiction:
Improvedetection capabilityVSAvoidspeed of reasoning
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

By segmenting the point cloud into local regions and processing them independently with a lightweight network, the patent reduces the computational burden per region and enables parallel processing, thereby improving detection speed while maintaining capability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs a lightweight, simple neural network that can be quickly deployed and executed on edge devices, replacing the need for large, resource-intensive transformer models. This lightweight model provides sufficient detection capability with much lower computational costs and faster execution.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Speed

If 4PCSs algorithm is used for pose estimation with less overlapping point clouds, then processing speed is improved, but detection accuracy deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidpose estimation accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The patent segments the point cloud into multiple local regions and processes each region separately to identify key points. This segmentation allows the algorithm to work effectively even when global point cloud overlap is minimal, as each local region can be processed independently to contribute to the overall pose estimation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary processing of point cloud regions to extract key local features and establish local coordinate systems before performing pose estimation. This preliminary action prepares the data in a form that enables accurate pose calculation even with limited overlap between reference and target point clouds.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4712022A1Method and apparatus for detecting target object, electronic device, and computer storage medium
Publication Date: 2026.03.18 SIEMENS AG
  • EP4712022A1 patent drawingFigure 1
  • EP4712022A1 patent drawingFigure 2
  • EP4712022A1 patent drawingFigure 3A~3B

AI summary

The present disclosure relates to a method and apparatus for detecting a target object, an electronic device, and a computer storage medium. The method includes: acquiring a target image and a first global point cloud of the target object; detecting, by using a detection model, first key regions; acquiring first regional point clouds corresponding to the first key regions; respectively calculating coordinates of respective midpoints in the first key regions; calculating a first transformation matrix based on coordinates of each of the midpoints and coordinates of reference midpoints; transforming, by using the first transformation matrix, the first global point cloud to obtain a second global point cloud; determining second key regions respectively corresponding to the first key regions; determining coordinates of respective midpoints and second sub point clouds in the second key regions; calculating a second transformation matrix based on reference point clouds and the second sub point clouds; calculating a third transformation matrix based on the second transformation matrix and the first transformation matrix; and determining pose information of a robot grabbing the target object based on reference pose information of a reference object and the third transformation matrix.