Closed-Loop Control of Additive Manufacturing via Reinforcement Learning
Additive manufacturing suffers from imperfections in hardware control and material consistency. As a result, the deposition of a large range of materials requires on-the-fly adjustment of process parameters. Unfortunately, learning the in-process control is challenging. The deposition parameters are complex and highly coupled, artifacts occur after long time horizons, available simulators lack predictive power, and learning on hardware is intractable. In this work, we demonstrate the feasibility of learning a closed-loop control policy for additive manufacturing. To achieve this goal, we assume that the perception of a deposition device is limited and can capture the process only qualitatively. We leverage this assumption to formulate an efficient numerical model that explicitly includes printing imperfections. We further show that in combination with reinforcement learning, our model can be used to discover control policies that outperform state-of-the-art controllers. Furthermore, the recovered policies have a minimal sim-to-real gap. We showcase this by implementing a first-of-its-kind self-correcting printer.
PDF Abstract
