RINGS: Collaborative Inference and Learning between Edge Swarms and the Cloud

University Of Texas At Austin, Austin TX

Investigators

Abstract

Fleets of networked robots are being deployed on our roads, in factories, and in hospitals for tasks like self-driving, manufacturing, and nurse assistance. These robots are struggling to process growing volumes of rich sensory data and deploy compute-and-power hungry machine learning (ML) models. However, robots have an opportunity to augment their intelligence by querying remote compute resources over next-generation (NextG) wireless networks. However, researchers lack algorithms to balance the accuracy benefits of networked computation with systems costs of delay, power, congestion, and load on remote compute servers. As such, this project is innovating algorithms based on decision theory (i.e., a mathematical cost-benefit analysis) to decide how to balance on-robot and remote computation while only communicating task-relevant, privacy-preserving data. The resulting communication-efficient algorithms aim to improve the resiliency of NextG networks by minimizing congestion. The project’s outreach efforts aim to enable K-12 students to prototype on remote robots. This project develops algorithms to enable joint inference, learning, and control between robotic swarms and the cloud while resiliently adapting to variations in network connectivity and compute availability. Today's robotic control algorithms are largely informed by onboard sensors and a local physical state, but effectively ignore the time-variant state of a network. As such, they often make sub-optimal decisions on when to query the cloud, often leading to excessive congestion. Accordingly, this project develops decision-theoretic algorithms that flexibly trade-off the accuracy benefits of the cloud with systems costs. First, the project develops collaborative inference algorithms that decide whether, and where, to offload computation using a Markov Decision Process. Then, it develops statistical data sampling algorithms that estimate the marginal gain of uploading new training data with labeling and training costs. The final thrust learns compressed representations of video and LiDAR that optimize for ML inference accuracy, as opposed to conventional human perception metrics. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.