Evaluating Distributed Computation Offloading Scalability for Multiple Robots

Perception in robotics using modern deep learning approaches is often computationally expensive. Cloud robotics and edge robotics provide possible solutions to this. In particular, computation offloading permits compute-constrained robots to offload compute tasks from the robot itself to more capable servers (the remote brain). However, the scalability of computation offloading for robots sharing a WiFi network has typically not been considered. In this work, we investigated the scalability of offloading to the cloud and edge for deployments of multiple robots in real-world settings and network conditions. We interpret typical network performance metrics such as latency, throughput, and packet loss in terms of their effect on robot computations to help to understand this question. To characterize the problem further, we introduce three different static offloading profiles based on the sensing capabilities of currently available robot platforms and examined these in a number of experiments on both simulated and physical wireless networks. Our results show that WiFi network capacity is much less than advertised when subjected to offloading data traffic and indicates limitations on the number of robots that can concurrently use computation offloading in a given space and the amount of data that they can transfer. This has significant implications for the dense deployment of robots that depend on computation offloading to meet their required service level.