Distributed Training and Management of Al Powered Robots using Teleoperation via Virtual Spaces

Technology

The training method proposed by the invention has two primary components. The first is a virtual “waiting room” where human operators can oversee a collection of robots and select one to manipulate. The second component is a partial algorithm for incorporating human input into the robot’s self-correction algorithm. The human operator and robot AI take on a “master-apprentice” dynamic, where the robot attempts to perform a task. If it fails a specified number of times or is not confident in the results the human operator is signaled. They can then take control of the robot to perform the task. The AI will learn from the way the operator performed the task and be able to replicate the effort unaided in the future. Thus, a single operator can be responsible for multiple robots by watching the “waiting room” and responding when an individual robot needs help with a task. 

Problem Addressed

Current algorithms and methods for training manufacturing AIs require a prolonged period of trial and error and may also take inordinate amounts of human time demonstrating successful and unsuccessful tasks until the robot can reliably perform the desired task. This method allows for more efficient training, by integrating AI self-feedback with human demonstrations when necessary. This results in higher task completion rates when compared to robots trained with SL methods and less human intervention when compared to robots trained with LFD methods. Additionally, a single human operator can use this method to train more AIs at the same time when compared to other available methods. 

Advantages

• Seamless incorporation of human feedback in manufacturing robot algorithms 

• More efficient than other methods of AI training 

• Higher task completion rates  

• Less human intervention in the training process 

Publications

Lipton, J. I., A. Schulz, A. Spielberg, L. Trueba, W. Matusik, and D. Rus. "Robot Assisted Carpentry for Mass Customization." In Proceedings of the IEEE International Conference on Robotics and Automation, 2018. Accessed February 14, 2022.