A learning system controls plasma shape despite sensor failures

Tokamak plasma shape control typically requires equilibrium reconstruction followed by fixed linear controllers, and breaks when sensors fail. This work trains a reinforcement learning agent in high-fidelity simulation on 120 experimental DIII-D plasma shapes, exposing it to random shape targets updated every 0.25 s to cover diverse transitions. The agent achieves 2.01 cm mean shape error on held-out configurations and robustly handles 30% random sensor dropout without switching logic or backup controllers. An asymmetric actor-critic architecture uses privileged equilibrium information to improve estimates under partial observability, while an auxiliary reconstruction head enables interpretability. The policy transfers zero-shot to experimental DIII-D shots and an independent simulator, directly commanding coil actuators on dynamic shape maneuvers.