Wrist-Mounted Vision for Robotic Grasp Pose Versatility
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robotic control systems are limited in their ability to execute a wide variety of grasp poses due to restricted camera views, typically only allowing top-down grasps, which restricts the number of possible grasp orientations and positions a robotic gripper can achieve.
Innovation Solution
A system employing a wrist-mounted camera oriented towards the gripper, utilizing deep reinforcement learning and a double deep Q-network to learn a mapping from images to estimated Q-values, allowing the robotic gripper to grasp objects from various angles and orientations by training in simulation and transferring to real-world scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a top-down camera view is used, then the system is simple to implement, but the number of possible grasp orientations and positions is limited
Solution Approach 1:
Instead of using a traditional top-down overhead camera view, the patent inverts the camera positioning to be wrist-mounted on the robotic arm, orienting the camera towards the gripper. This inversion allows the camera to capture images from the gripper's perspective, enabling the system to determine grasp poses for objects viewed from side, angled, and other non-top-down orientations, thereby significantly increasing the number of achievable grasp poses.
2Adaptability or versatility
If a wrist-mounted camera is used, then more grasp poses are achievable, but the system complexity increases
Solution Approach 1:
The wrist-mounted camera serves multiple functions: it captures images of objects from the gripper's perspective for pose estimation, provides visual feedback for grasp execution, and enables the system to handle various object types and orientations. This multi-functionality justifies the increased device complexity by delivering comprehensive grasp capabilities across diverse scenarios.
3Adaptability or versatility
If deep reinforcement learning is used for grasp planning, then grasp versatility improves, but computational resources required increase
Solution Approach 1:
The system performs preliminary action by training the deep reinforcement learning model offline using simulation data before actual grasp execution. During real-world operation, the pre-trained model quickly processes images and determines grasp poses without requiring extensive real-time computation. This separates the computationally intensive training phase from the efficient inference phase, reducing real-time energy consumption while maintaining high grasp versatility.
Data Source
AI summary
In at least one embodiment, under the control of a robotic control system, a gripper on a robot is positioned to grasp a 3-dimensional object. In at least one embodiment, the relative position of the object and the gripper is determined, at least in part, by using a camera mounted on the gripper.