Tactile Grasp Policy Learning for Unseen Object Shapes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Robotic systems face challenges in autonomously grasping objects of varying shapes and sizes with multi-fingered hands, often requiring significant computational time and being sensitive to perception and calibration errors, which can lead to unsuccessful grasping.
Innovation Solution
A deep reinforcement learning approach using tactile sensing and human grasping demonstrations to learn a policy that generalizes across object geometries, leveraging context variables and tactile feedback to adapt to new shapes without further training, bridging the simulation-to-real-world gap.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional robotic control systems are used to determine grasp positions and orientations, then the system can execute grasps on objects, but the computational time required is significant and the system is sensitive to perception and calibration errors
Solution Approach 1:
The system pre-trains neural network policies in simulation environments with diverse object geometries before deployment. This preliminary training allows the robot to quickly adapt to new objects without extensive computational optimization during actual operation, reducing real-time computational time while maintaining high grasp success rates
Solution Approach 2:
Traditional model-based grasp planning algorithms are replaced with data-driven neural network policies. This substitution eliminates complex computational geometry processing and optimization routines, significantly reducing computational time while improving robustness to perception errors through the neural network's inherent error tolerance
2Adaptability or versatility
If traditional robotic control systems are used to grasp objects, then grasps can be executed, but the system requires significant computational resources and is sensitive to perception errors
Solution Approach 1:
A single neural network policy is trained to handle multiple object geometries and grasp types simultaneously. The policy learns universal grasp strategies that generalize across diverse objects, eliminating the need for separate specialized algorithms for different object classes and reducing overall system complexity
Solution Approach 2:
The system uses simulation-to-real transfer by creating virtual copies of physical objects in training environments. Neural networks are trained on these simulated copies with varied geometries, allowing the system to learn from extensive virtual experience without proportional increases in physical hardware complexity
3Reliability
If model-based grasp planning is used, then grasps can be planned and executed, but the system requires significant computational time and is sensitive to calibration errors
Solution Approach 1:
Complex model-based planning algorithms involving geometric reasoning and constraint satisfaction are replaced with neural network inference. The neural network directly maps object features to grasp parameters, eliminating multi-step computational processes while improving robustness to calibration errors through learned invariance
Data Source
AI summary
Apparatuses, systems, and techniques to perform a grasp of on object using an articulated robotic hand equipped with one or more tactile sensors. In at least one embodiment, a machine-learned model trained in simulation to grasp a cuboid using signals received from tactile sensors is applied to grasping objects of various shapes in a real-world environment.