6-DOF Grasp Planning With Collision Checking in Clutter
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Robotic systems face challenges in determining stable grasp poses for objects in cluttered environments, especially when complete 3D models are not available, leading to collisions and reduced grasp success rates.
Innovation Solution
A learning-based approach that generates 6-DOF grasps using partial point cloud observations, incorporating instance segmentation and a cascaded grasp generation method to reason about object-level grasps and check for collisions, allowing for effective grasp sequences in cluttered scenes without requiring complete object models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If geometry-inspired heuristics are used to select grasp points, then grasp selection efficiency is improved, but reliability deteriorates due to dependence on complete 3D models which are unavailable in noisy camera scenarios
Solution Approach 1:
The system performs preliminary action by generating a complete 3D model of the target object using RANSAC-based plane fitting and point cloud processing before grasp selection. This pre-established model provides reliable geometric information even when initial camera observations are noisy or incomplete, enabling subsequent grasp algorithms to operate on accurate data rather than raw sensor inputs
Solution Approach 2:
The patent introduces an intermediary component - a learned collision checker neural network - that mediates between the grasp selection process and the cluttered environment. This intermediary evaluates potential grasps by predicting collisions with surrounding objects, allowing the system to select reliable grasps without requiring complete knowledge of all objects in the scene
2Measurement precision
If complete 3D models are required for accurate grasp determination, then grasp accuracy is improved, but device complexity increases and performance deteriorates in realistic noisy scenarios
Solution Approach 1:
The system applies partial action by generating 3D models and performing collision checks only for the specific target object and relevant surrounding objects, rather than processing the entire scene. The RANSAC algorithm fits planes to point clouds selectively, and the collision checker focuses on predicting collisions for candidate grasps, avoiding unnecessary computation on irrelevant data
Solution Approach 2:
The patent replaces traditional mechanical/geometric collision checking methods with a learned collision checker neural network. Instead of using complex geometric algorithms to compute intersections between gripper trajectories and objects, the system uses a trained neural network that predicts collisions based on image and point cloud inputs, significantly reducing computational complexity
3Measurement precision
If traditional collision checking methods are used, then measurement precision is maintained, but productivity decreases due to computational overhead in cluttered scenes
Solution Approach 1:
The system substitutes traditional geometric collision checking with a learned collision checker neural network. The neural network is trained to predict whether a grasp will collide with surrounding objects, replacing computationally intensive geometric intersection algorithms with a faster inference-based approach that maintains accuracy while improving speed
Data Source
AI summary
Apparatuses, systems, and techniques determine a set of grasp poses that would allow a robot to successfully grasp an object that is proximate to at least one additional object. In at least one embodiment, the set of grasp poses is modified based on a determination that at least one of the grasp poses in the set of grasp poses would interfere with at least one additional object that is proximate to the object.