Robotic Grasp Planning With Robustness Prediction for Unseen Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Robots face challenges in reliably grasping a variety of previously unseen objects due to inaccurate sensors and limited precision in gripper control, making consistent robust grasping difficult, especially in environments with unique or difficult-to-grasp items.
Innovation Solution
A computer-implemented method using 3D object models, analytic mechanical representations, and statistical sampling to generate a training dataset for training a function approximator that selects robust grasp configurations for a robotic grasping mechanism, combining machine learning with physics-based models to predict grasp quality and robustness, and leveraging deep learning with large datasets for efficient grasp planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sensors and grippers are used for object grasping, then object manipulation capability is enabled, but measurement precision and control precision are insufficient leading to unreliable grasping
Solution Approach 1:
The system performs preliminary actions by generating multiple candidate grasp configurations and evaluating their robustness before actual grasping. The robustness evaluation predicts potential grasp failures in advance, allowing the system to select or adjust grasp plans before execution, thereby compensating for sensor inaccuracies and control limitations.
Solution Approach 2:
The system changes parameters by adjusting grasp configuration parameters (such as grasp pose, contact points, and grip force distribution) based on robustness evaluation results. This allows the system to optimize grasping reliability by selecting configurations that are most tolerant to sensor errors and control imprecision.
2Reliability
If machine learning is used to train grasp prediction systems, then grasp prediction capability is improved, but training data requirements become impractically large
Solution Approach 1:
The system uses copying by creating synthetic training data through simulation environments that replicate real-world grasping scenarios. Instead of requiring millions of physical robot trials, the system generates virtual grasp examples with known outcomes, which can be used to train machine learning models efficiently without extensive physical data collection.
Solution Approach 2:
The system performs preliminary analysis by using physics-based models and analytical methods to pre-evaluate grasp robustness and generate labeled training data. This preliminary computational work reduces the need for extensive machine learning training by providing high-quality synthetic datasets that capture the essential characteristics of robust grasping.
3Adaptability or versatility
If specialized grasping hardware is used to compensate for unknown object properties, then grasping capability for unseen objects is improved, but visibility of the object in the gripper is limited
Solution Approach 1:
The system achieves universality by developing a general-purpose grasping system that can handle diverse object types without specialized hardware. The robustness evaluation framework and machine learning model are designed to work with various object geometries and properties, making the system adaptable to unseen objects while maintaining full visual observation capability.
Solution Approach 2:
The system uses an intermediary computational model (robustness evaluation function) that mediates between sensor data and grasping actions. This intermediary layer processes sensor information to predict grasp outcomes and guide gripper positioning, enabling the system to grasp unseen objects reliably without needing specialized hardware that would block visibility.
Data Source
AI summary
Embodiments are generally directed to generating a training dataset of labelled examples of sensor images and grasp configurations using a set of three-dimensional (3D) models of objects, one or more analytic mechanical representations of either or both of grasp forces and grasp torques, and statistical sampling to model uncertainty in either or both sensing and control. Embodiments can also include using the training dataset to train a function approximator that takes as input a sensor image and returns data that is used to select grasp configurations for a robot grasping or targeting mechanism.