Force-Torque Robotic Assembly Control Without Vision Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional robotic control techniques struggle to adapt to unstructured environments, such as construction sites, due to the difficulty in calibrating motion capture or vision-based systems, and they are often robot-specific, limiting their applicability to diverse and unpredictable conditions.
Innovation Solution
A machine learning model trained via reinforcement learning that uses force and torque data to control robots, enabling robot-agnostic policy deployment across various robotic platforms without requiring pose data from motion capture or vision-based systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If motion capture or vision-based systems are used to capture pose data, then robot control accuracy is improved, but system complexity and calibration difficulty increase
Solution Approach 1:
The patent extracts and removes the complex motion capture and vision-based pose estimation systems from the robotic control loop. Instead of using these external sensing systems, the invention directly utilizes force and torque sensor data from the robot's end effector to guide control decisions, eliminating the need for complex calibration and setup while maintaining effective control in unstructured environments
Solution Approach 2:
The patent replaces optical/mechanical sensing systems (motion capture cameras, vision systems) with a purely force-based sensing approach. By substituting mechanical/optical measurement systems with force sensor measurements and machine learning-based inference, the system achieves pose-guided control without the complexity of traditional sensing infrastructure
2Reliability
If vision-based systems are used to infer pose data, then robot control is improved, but performance deteriorates in contact-rich phases with occlusion and poor lighting
Solution Approach 1:
The patent introduces force and torque sensors as intermediary measurements that indirectly provide information about robot pose and environment interaction without requiring direct visual observation. These tactile measurements serve as a mediator that is immune to occlusion and lighting conditions, allowing the machine learning model to infer pose information through force-based sensing rather than vision
Solution Approach 2:
The patent replaces vision-based pose estimation with force-based sensing and machine learning inference. By substituting the optical measurement system with mechanical force sensors and computational models, the system achieves reliable control in contact-rich phases where vision systems fail due to occlusion and poor lighting
Data Source
AI summary
Techniques are disclosed for training and applying machine learning models to control robotic assembly. In some embodiments, force and torque measurements are input into a machine learning model that includes a memory layer that introduces recurrency. The machine learning model is trained, via reinforcement learning in a robot-agnostic environment, to generate actions for achieving an assembly task given the force and torque measurements. During training, experiences are collected as transitions within episodes, the transitions are grouped into sequences, and the last two sequences of each episode have a variable overlap. The collected transitions are stored in a prioritized sequence replay buffer, from which a learner samples sequences to learn from based on transition and sequence priorities. Once trained, the machine learning model can be deployed to control various types of robots to perform the assembly task based on force and torque measurements acquired by sensors of those robots.