Human remote monitoring of robotic fleets is commonly used in applications such as self-driving taxis or implement automatic repositories.
Industrial robots in the Mercedes-Benz American International Plant located in Tuscaloosa County. Original image from the collection of Carol M. Highsmith’s America, Library of Congress. Digital enhancement by rawpixelCC0 . public domain
Here, any individual robot can share its intervention data with the rest of the fleet i.e. a set of independent robots that simultaneously implement the same control policy for same task in parallel environments. A fundamental issue in this mission is how to allocate limited human supervision to the robot to maximize fleet throughput.
A recent article on arXiv.org covers IFL Benchmark (Interactive Fleet Learning), a new open source Python toolkit and benchmark for developing and evaluating human-to-robot attribution algorithms for learning by group. The researchers also propose a new algorithm for the IFL mission, which significantly affects the robot fleet’s performance. Not only does it learn where to allocate people, but it also knows when to stop asking for unnecessary supervision.
Commercial and industrial deployments of robotic fleets often fall into the realm of human remote control during execution when the robot is at risk or unable to keep up with the task schedule. With continuous learning, remote team interventions can also be used to improve robotic fleet control over time. A central question is how to effectively allocate limited human attention to individual robots. Previous work addressed this in a one-man, one-robot installation. We formalize the Interactive Group Learning (IFL) setting, in which multiple robots interactively query and learn from multiple human supervisors. We present the fully implemented open source IFL benchmark suite of the GPU-accelerated Isaac Gym environment for the evaluation of IFL algorithms. We propose Fleet-DAgger, a family of IFL algorithms, and compare a new Fleet-DAgger algorithm with 4 baselines in the simulation. We also performed 1000 times of physical block push experiments with 4 ABB YuMi robotic arms. Tests show that allocating humans to robots significantly affects robot team performance, and our algorithm achieves up to 8.8x better return on human effort than baseline. department. See This https URL for code, video and additional documentation.
Research articles: Hoque, R., “Fleet-DAgger: Fleet of Interactive Robots Scalable Human-Supervised Learning”, 2022. Link: https://arxiv.org/abs/2206.14349
