Embodied AI and Data Generation for Manufacturing Robotics

Summary

Approaches categorized as “artificial intelligence” (AI) are enabling significant advances in robotics. These include symbolic logic, Bayesian statistics, and numerous other algorithmic approaches. Recently, data-centric machine learning has become a prominent tool in a number of disciplines relevant to robotics.  AI applied to robotics “can create smarter, faster, cheaper, and more environmentally-friendly production processes that can increase worker productivity, improve product quality, lower costs, and improve worker health and safety. Machine learning algorithms can improve the scheduling of manufacturing processes and reduce inventory requirements.”  AI’s rapid rate of adoption has led to many successes, as well as the need for a measurement science infrastructure to help generate data and qualify it.  For industry to use AI, they must trust what comes out of the AI system. The key idea for this project is to develop data sets and trained AI system, validated through performance evaluation techniques, to allow them to be applied to manufacturing robotic systems. This will allow manufacturers to gain more value from their robots by allowing the robot to “learn” new tasks, and how to better perform existing tasks, without the need for human intervention. NIST is uniquely qualified to address this because of our experience in robot performance characterization, information modeling standards, and robot programming.

Description

To provide structured artificial intelligence (AI) and machine learning (ML) training datasets, and proven, trained, and applied AI/ML models to improve the performance and autonomy of manufacturing robotic applications.

What is the Technical Idea?

In this project, we will use the term AI to mean both the traditional approaches as well as machine learning (ML). The project will develop data generating mechanisms (physical and virtual) across a variety of relevant experimental factors relevant to industrial robotic applications.  This could include generating training data from simulation tools, gathering training data from web resources, gathering data from multiple physical trials, among other approaches. Data generating mechanism could include The generated data will be organized and formatted in a way that will facilitate the training of AI applications in robotic software agility, perception, grasping, teaming, and mobility, and will include metadata that will define the data’s scope and applicability. The datasets and the associated approaches will be validated through its implementation on an embodied agent (a robot), showing how the embodied agent can directly solve a manufacturing task or improve existing performance.
The generated data will be made available to researchers and industry practitioners to train and test AI systems for industrial robotic applications. The details regarding learning methods, settings, and code bases will be disseminated through publications. Furthermore, trained AI system models will be made publicly available.

What is the Research Plan?

The project will take an iterative approach throughout its three-year duration, focusing on generation of AI training data, dissemination of AI training data, and development and dissemination of AI/ML models.

Initial efforts will focus on designing an approach for the generation and curation of AI training data (subsequently referred to as data generation), scoped for selected thrust areas, based on a review of relevant AI/ML strategies and leveraging public AI/ML codebases. Throughout the course of this project, the data generation mechanisms will be hardened, and the resulting AI/ML training datasets will be released to the public. These datasets will focus on specific manufacturing tasks (e.g., peg-in-hole, kitting), or capabilities (e.g., perception, grasping, and path planning).  

The dataset will be used within the project to develop and train AI/ML models. These models will be applied to relevant manufacturing challenges to show their value. Once validated, the models will be documented and released to the public.