A decoupled perception-control training pipeline for vision-based deep reinforcement learning, with cross-domain and...

Training a control policy directly from camera pixels with deep reinforcement learning is fragile: the visual encoder and the policy are optimised by the same temporal-difference gradient, so the latent state is non-stationary, and recovering a working agent usually demands image augmentation, a separate encoder learning rate and a tuned replay ratio. This article specifies the decoupled alternative as a complete, replicable pipeline, in which a convolutional autoencoder is pre-trained on a small demonstration corpus, frozen, and then used as a fixed feature map for an off-policy actor-critic agent. The pipeline customises two established methods - convolutional autoencoder representation learning and Soft Actor-Critic - and its contribution is the set of specifications that make the combination work with no per-configuration tuning, together with stage-wise validation checkpoints that expose a failed perception stage before expensive policy training begins. Two extensions reuse the same artifacts, one for deployment in an unseen visual domain and one for a revised objective, each with the algorithm substitution and the evaluation rules it requires. Under a single hyperparameter set applied unchanged, every decoupled configuration converged, while an end-to-end baseline consuming the same 120 × 160 × 3 observation never completed a lap.