Tactile Pose Estimation for Tight-Tolerance Connector Assembly
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robotic assembly methods struggle with accurately estimating the 3D pose of mating components for tight tolerance assemblies, especially with small parts, and face challenges in recovering from pose misalignment and contact formations, particularly in complex geometries.
Innovation Solution
A system using image-based tactile sensors at the gripper fingers for high-accuracy pose estimation, combined with a simple impedance controller, to refine the pose estimation through a two-phase process involving matching observed tactile depth images to a precomputed set and refining with ICP, enabling precise robotic assembly of complex connectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If visual 3D pose estimation is used for assembly, then the system is simple to implement, but the measurement precision is insufficient for tight tolerance assemblies
Solution Approach 1:
The patent combines visual sensors and tactile sensors into an integrated sensing system. The visual sensor captures images of the object, while the tactile sensor detects contact forces and torques. By merging these two sensing modalities, the system achieves high-precision pose estimation that leverages both optical information and mechanical contact information, resolving the contradiction between measurement precision and device complexity.
Solution Approach 2:
The patent introduces a neural network as an intermediary that processes visual sensor data and predicts tactile sensor readings. This intermediary model enables the system to infer pose information from visual data alone, achieving high measurement precision without requiring complex real-time tactile sensing, thus resolving the contradiction between precision and complexity.
2Adaptability or versatility
If search patterns with compliance are used for insertion, then simple geometries can be assembled, but the method does not generalize to complicated geometries
Solution Approach 1:
The patent employs dynamic impedance control that adapts the robot's mechanical impedance during insertion based on real-time sensory feedback. The controller dynamically adjusts stiffness and damping parameters to handle varying contact conditions, enabling successful insertion of complex geometries while maintaining ease of operation through automated adaptive control.
Solution Approach 2:
The patent implements a closed-loop feedback system where tactile sensor measurements of contact forces and torques are continuously fed back to the controller. This feedback enables real-time adjustment of insertion trajectories and forces, allowing the system to handle complicated geometries adaptively while keeping the operation simple through automated control.
3Adaptability or versatility
If deep reinforcement learning is used for complex geometries, then adaptability improves, but sample efficiency and generalization deteriorate
Solution Approach 1:
The patent pre-trains neural network models offline using simulated tactile data before deployment. This preliminary action prepares the system to handle complex geometries by learning from diverse virtual examples, enabling fast adaptation to real-world scenarios without requiring extensive online training, thus reducing time loss while maintaining versatility.
4Measurement precision
If tactile sensors are used for pose estimation, then measurement precision improves, but device complexity increases
Solution Approach 1:
The patent designs the tactile sensor to serve multiple functions: it detects contact forces, measures torques, and provides pose estimation information. This multi-functionality allows a single sensor component to achieve high measurement precision while minimizing the increase in device complexity by avoiding the need for separate sensors for each measurement type.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A pose controller is provided for controlling a pose of an object to assemble with a mating object by a gripper of a robot arm. The pose controller includes an interface configured to receive tactile signals from the tactile sensors and transmit a control signal to the actuators, a processor, and a memory, in association with the processor, configured to store a precomputed set of tactile depth images and instructions of computer-implemented method. The instructions cause the processor to perform steps of computing measured tactile depth images from the received tactile signals, refining the pose of the object by matching between the precomputed set of tactile depth images and the measured tactile depth images by a point-to-plane iterative closest point (ICP) algorithm, generating a gripper trajectory command based on the refined pose of the object and a target pose of the object, wherein the target pose is aligned against the mating object with a nominal distance above the mating object, and controlling the actuators of the robot arm according to a gripper trajectory by transmitting the gripper trajectory command to the robot controller.