Collaborative Perception And Learning Between Robots And The Cloud

Augmenting robotic intelligence with cloud connectivity is considered one of the most promising solutions to cope with growing volumes of rich robotic sensory data and increasingly complex perception and decision-making tasks. While the benefits of cloud robotics have been envisioned long before, we have historically lacked flexible methods to trade-off the benefits of cloud computing with end-to-end systems costs of network delay, cloud storage, human annotation time, and cloud-computing time. To address this need, this thesis introduces decision-theoretic algorithms that allow robots to significantly transcend their on-board perception capabilities by using cloud computing, but in a low-cost, fault-tolerant manner. The utility of these algorithms is demonstrated on months of field data and experiments on state-of-the-art embedded deep learning hardware. Specifically, for compute-and-power-limited robots, this thesis presents a lightweight model selection algorithm that learns when a robot should exploit low-latency on-board computation, or, when highly uncertain, query a more accurate cloud model. Then, I present a collaborative learning algorithm that allows a diversity of robots to mine their real-time sensory streams for valuable training examples to send to the cloud for model improvement. This thesis concludes by surveying a number of future research directions on the systems and theoretical aspects of networked system control, some of which extend beyond cloud robotics.