End-to-end autonomous steering models map raw sensory observations directly to continuous control actions, reducing the hand-engineering burden of modular autonomy stacks. However, two practical barriers limit deployability: (i) steering prediction is inherently temporal and frame-wise convolutional neural networks (CNNs) often exhibit oscillations and degraded recovery under covariate shift, and (ii) centralized training requires aggregating large-scale driving data that are distributed, bandwidth-intensive, and privacy-sensitive. This paper presents a federated temporal autonomous steering framework combining CNN-based spatial perception with Long Short-Term Memory (LSTM) sequence modeling under Federated Learning. Multiple clients (simulated vehicles) train locally on heterogeneous, non-identically distributed (non-IID) driving data and periodically synchronize with a server via Federated Averaging. We formalize the temporal imitation learning objective, present a sequence-aware model architecture and preprocessing pipeline, and define a federated optimization protocol with communication and non-IID considerations. Extensive experiments compare centralized CNN, centralized CNN-LSTM, and federated CNN-LSTM baselines under steering imbalance and distribution shifts induced by curvature bias and lighting perturbations. Results show that temporal modeling reduces steering oscillation variance and improves closed-loop stability, while federated training achieves comparable accuracy to centralized training with privacy-preserving data locality. We also analyze convergence behavior under heterogeneous clients, communication cost per round, and ablations on sequence length, local epochs, and client sampling. The proposed framework provides a reproducible foundation for distributed, privacy-preserving end-to-end control.