Point Cloud Grasp Planning for Unseen Objects in Clutter
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional bin picking methods in dense clutter environments are domain-dependent and fail to perform well when encountering diverse, unseen objects with varying colors, textures, shapes, and sizes, due to sensor noise and uncertainty in robot actuation, requiring extensive training data and being sensitive to equipment changes.
Innovation Solution
A point cloud-based grasp planning framework that generates sampled grasp poses, computes depth difference values, creates binary maps, identifies subregions, selects feasible poses, refines them by shifting and adjusting, and uses a Grasp Quality Score to determine optimal grasp poses for robotic arms, independent of specific domains and equipment configurations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional bin picking methods are used in dense clutter environments, then grasp planning can be performed, but the methods are domain-dependent and fail to perform well on diverse unseen objects
Solution Approach 1:
The patent uses depth maps as copies of the physical scene to perform grasp planning. Instead of directly interacting with diverse physical objects, the system creates and processes depth map representations, which can be universally applied across different object types without domain-specific training. This copying approach enables domain-independent grasp planning while maintaining reliability.
2Measurement precision
If deep learning-based methods are used for grasp planning, then grasp quality can be predicted, but extensive training data is required which is time-consuming to collect
Solution Approach 1:
The system performs self-service by generating its own training data through simulation. Instead of requiring external collection of thousands of real-world grasp samples, the method uses simulated depth maps and corresponding grasp outcomes to train the neural network. This self-generated dataset approach eliminates time-consuming real-world data collection while achieving accurate grasp quality prediction.
3Loss of time
If CNN models are trained on simulated depth scan datasets, then training time is reduced, but the models become sensitive to parameters such as robot gripper, depth camera, and camera-to-workspace distance
Solution Approach 1:
The patent creates a universal grasp planning framework that works across different equipment configurations. By training on diverse simulated scenarios with varying camera positions, gripper types, and scene configurations, the neural network learns parameter-invariant features. This universal training approach allows the same model to perform reliably across different robot grippers, depth cameras, and workspace distances without retraining.
4Productivity
If domain-dependent methods are used for grasp planning, then performance can be optimized for specific domains, but the methods fail when the target domain differs from the source domain
Solution Approach 1:
The system creates depth map copies of various physical scenes and trains on these representations rather than domain-specific features. This copying approach allows the model to learn generalizable grasp planning skills from simulated environments that can be transformed into any target domain through depth map generation, achieving both high productivity and cross-domain adaptability.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A fully automated and reliable picking of a diverse range of unseen objects in clutter is a challenging problem. The present disclosure provides an optimum grasp pose selection to pick an object from a bin. Initially, the system receives an input image pertaining to a surface. Further, a plurality of sampled grasp poses are generated in a random configuration. Further, a depth difference value is computed for each of a plurality of pixels corresponding to each of the plurality of sampled grasp poses. Further, a binary map is generated for each of the plurality of sampled grasp poses and a plurality of subregions are obtained. Further, a plurality of feasible grasp poses are selected based on the plurality of subregions and a plurality of conditions. Further, the plurality of feasible grasp poses are refined and an optimum grasp pose is obtained based on a Grasp Quality Score (GQS).