Co-Design of Communication and Machine Inference for Cloud Robotics

Abstract

Today, even the most compute-and-power constrained robots can measure complex, high data-rate video and LIDAR sensory streams. Often, such robots, ranging from lowpower drones to space and subterranean rovers, need to transmit high-bitrate sensory data to a remote compute server if they are uncertain or cannot scalably run complex perception or mapping tasks locally. However, today's representations for sensory data are mostly designed for human, not robotic, perception and thus often waste precious compute or wireless network resources to transmit unimportant parts of a scene that are unnecessary for a high-level robotic task. This paper presents an algorithm to learn task-relevant representations of sensory data that are codesigned with a pre-trained robotic perception model's ultimate objective. Our algorithm aggressively compresses robotic sensory data by up to 11 × more than competing methods. Further, it achieves high accuracy and robust generalization on diverse tasks including Mars terrain classification with low-power deep learning accelerators, neural motion planning, and environmental timeseries classification.

Data Sharing and Compression for Cooperative Networked Control

“…A similar analysis for a linear encoder-decoder structure for networked inference, not control, is presented in [13]. The key difference from our current paper is our problem setup is for control, not networked inference.…”

“…The cost function (Eq. 13) has γ e = γ s = γ u = 1 to equally penalize deviation from the set-point and regulation effort. Finally, we collected two weeks of stochastic timeseries of temperature, pressure, humidity, and light from the Google Edge Tensor Processing Unit (TPU)'s environmental sensor board for our experiments, as detailed in the supplement.…”

“…While the term co-design appears in select work on networked LQR, it refers to a drastically different setting where a communication scheduler and tele-operated controller must be jointly designed [8][9][10][11]. Finally, our work differs from deep neural network (DNN) compression schemes for video inference [12,13] since we focus on control.…”

“…Fig.12, Fig 13. and Fig.14present the time-domain forecasts with different bottleneck dimensions Z for IoT, taxi scheduling, and battery charging scenarios, respectively.…”

Performance of Sensor Data Process Offloading on 5G-Enabled UAVs

“…However, offloading relies on the transmission of the sensor data to be processed to the considered cloud. Often, such sensor data transmission requires extensive communication resources and is better suited for human perception rather than computer processing [ 11 ]. Similar scenarios are present in different fields, such as smartphone cloud gaming [ 12 ] or streaming HD video applications to the cloud for extensive computer vision processing [ 13 , 14 ].…”