Reinforcement Learning for Energy-Efficient Cloud Robotics

Cloud robotics offloads compute-heavy tasks — such as deep-network inference for semantic segmentation — from lightweight, compute-constrained robots to a cloud server, saving local compute and battery at the cost of network congestion and potential delay for time-critical tasks like autonomous driving.

This thesis learns an RL-based policy for deciding when to offload computation to the cloud versus process it locally, balancing accuracy, network latency, and battery life across one or more inference models. Alternative tasks can be discussed with the supervisor before starting.