Robotic Grasp Learning Using Simulated Bin-Picking Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot grasp generation methods are computationally expensive, require manual teaching, and are not optimal for real-world scenarios involving jumbled parts and collision avoidance, limiting their adaptability to unknown objects and occlusions.
Innovation Solution
A method that uses a database of solid and surface models for objects and grippers, with random initialization and iterative optimization to compute diverse grasps, followed by physical environment simulation to train neural networks for real-world robotic operations, correlating simulation results with camera depth image data to identify high-quality grasps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If learning-based grasp detection methods are used to adapt to unknown objects, then adaptability to unknown objects and occlusions is improved, but computational cost increases and manual teaching requirements arise
Solution Approach 1:
The patent pre-computes and stores high-quality grasp candidates in a database before actual robotic operations. The system performs iterative optimization offline to generate comprehensive grasp data for various objects and grippers, which is then quickly retrieved and used during real-time operations, eliminating the need for complex computational processes during actual task execution.
Solution Approach 2:
The patent creates a simulated physical environment that replicates real-world bin scenarios with jumbled parts. By training neural networks in this virtual environment using pre-computed grasp data, the system produces realistic grasp scenarios without requiring expensive physical trials, effectively copying real-world conditions in a computationally efficient virtual space.
2Reliability
If empirical trials are used to collect realistic grasp data, then realism of grasp scenarios is improved, but time consumption and resource requirements increase significantly
Solution Approach 1:
The patent creates a simulated physical environment that replicates real-world bin scenarios with jumbled parts. By training neural networks in this virtual environment using pre-computed grasp data, the system produces realistic grasp scenarios without requiring expensive physical trials, effectively copying real-world conditions in a computationally efficient virtual space.
Solution Approach 2:
The patent replaces physical empirical trials with computational simulation. Instead of physically manipulating real objects and grippers to collect grasp data, the system uses iterative optimization algorithms to compute grasp candidates and virtual environment simulation to evaluate them, substituting mechanical experimentation with computational modeling.
3Measurement precision
If traditional grasp generation methods are used with manual teaching, then control precision is improved, but ease of operation deteriorates due to significant time spent on heuristics design
Solution Approach 1:
The patent implements self-service through automated iterative optimization that generates high-quality grasp candidates without human intervention. The system automatically selects objects and grippers from databases, computes optimal grasps through iterative refinement, and evaluates them in simulated environments, eliminating the need for manual heuristic design while maintaining high grasp quality.
Solution Approach 2:
The patent transforms the grasp generation process from manual parameter specification to automated iterative optimization. Instead of manually defining grasp parameters and heuristics, the system automatically adjusts and refines grasp parameters through iterative computation, changing the approach from human-guided parameter setting to algorithm-driven parameter optimization.
Data Source
AI summary
A grasp generation technique for robotic pick-up of parts. A database of solid or surface models is provided for all objects and grippers which are to be evaluated. A gripper is selected and a random initialization is performed, where random objects and poses are selected from the object database. An iterative optimization computation is then performed, where many hundreds of grasps are computed for each part with surface contact between the part and the gripper, and sampling for grasp diversity and global optimization. Finally, a physical environment simulation is performed, where the grasps for each part are mapped to simulated piles of objects in a bin scenario. The grasp points and approach directions from the physical environment simulation are then used to train neural networks for grasp learning in real-world robotic operations, where the simulation results are correlated to camera depth image data to identify a high quality grasp.