Pixelwise Grasp Prediction Without Full 3D Object Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing robotic grasp generation methods rely heavily on complete 3D models of objects, which are not always available, limiting their effectiveness in real-world scenarios, and often lack efficiency in grasp determination.

Innovation Solution

A neural network-based grasping model generates pixelwise predictions for grasp outcomes, incorporating reinforcement learning to tune weights and combine with heuristic data, enabling robust grasp pose generation without relying on complete 3D models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If geometry-inspired heuristics and geometric analysis are used to select grasp points, then grasp stability and reachability can be analyzed, but complete 3D models of objects are required which are not always available

Engineering Contradiction:
Improvegrasp stabilityVSAvoid3D model requirement
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent uses 2D image data as a copy or representation of the 3D object, eliminating the need for complete 3D models. The pixelwise predictions are generated directly from 2D images, allowing grasp point selection without requiring full 3D reconstruction of the object geometry.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces traditional geometric and mechanical analysis methods with a neural network-based pixelwise prediction approach. Instead of using geometry-inspired heuristics that require 3D models, the system uses learned patterns from image data to directly predict grasp outcomes at the pixel level.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If traditional grasp generation methods using complete 3D models are used, then accurate grasp analysis can be performed, but the method efficiency and speed are reduced

Engineering Contradiction:
Improvegrasp determination accuracyVSAvoidgrasp generation speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the grasp determination task into pixelwise predictions, where each pixel independently contributes to the grasp outcome assessment. This segmentation allows parallel processing of all pixels simultaneously, significantly improving computational efficiency while maintaining accuracy through aggregate predictions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The neural network performs preliminary learning during training to capture complex grasp patterns and relationships. During inference, the pre-trained model can rapidly generate pixelwise predictions without performing computationally intensive geometric analyses, thus improving real-time grasp generation speed.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If pixelwise predictions are generated for multiple grasp outcomes, then comprehensive grasp assessment can be achieved, but the complexity of aggregating and processing multiple predictions increases

Engineering Contradiction:
Improvegrasp outcome assessmentVSAvoidprediction aggregation complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent merges multiple pixelwise predictions for different grasp outcomes into a unified grasp assessment. The neural network integrates predictions across multiple outcomes at the pixel level, allowing comprehensive evaluation of grasp quality, stability, and success probability through a cohesive aggregation process.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4255691B1Pixelwise predictions for grasp generation
Publication Date: 2026.02.04 OCADO INNOVATION LTD
  • EP4255691B1 patent drawingFigure 1
  • EP4255691B1 patent drawingFigure 2
  • EP4255691B1 patent drawingFigure 3A

AI summary

Provided are an apparatus and method for grasp generation, involving obtaining image data, comprising depth data, representative of an image of an object captured by a camera, and providing the image data to a grasping model comprising a neural network trained to predict a plurality of outcomes associated with a grasping operation independently based on images of graspable objects. A plurality of pixelwise predictions corresponding to the plurality of outcomes are obtained, each pixelwise prediction being a representation of pixelwise probability values corresponding to a given outcome of the plurality of outcomes associated with the grasping operation. The plurality of pixelwise predictions are aggregated to obtain an aggregated pixelwise prediction which is output for selection therefrom of one or more pixels on which to base generation of one or more grasp poses to grasp the object.