Real-Time Neural Model Predictive Control for Quadrotors via Input-Convex Temporal Convolutional Networks

Accurate dynamics models are essential for improving the tracking performance of model predictive control (MPC) for unmanned aerial vehicles (UAV). However, UAV dynamics are highly nonlinear, strongly coupled, and affected by aerodynamic effects with temporal dependence, which are difficult to capture using nominal rigid-body models alone. For this reason, temporal neural networks (TCNs) have been widely used to learn residual dynamics. Nevertheless, repeated evaluation of network outputs and sensitivities during online optimization incurs substantial computational overhead, hindering real-time implementation. Moreover, integrating temporal residual models into real-time MPC requires causally consistent feature histories at future shooting nodes and efficient horizonwise differentiation. To address these issues, this paper proposes a real-time neural MPC framework that integrates an input-convex TCN residual model, history-conditionedbatched local linearization, and real-time iteration sequential quadratic programming (RTI-SQP).The proposed input-convex temporal convolutional network (IC-TCN) employs causal dilated convolutions and sign-symmetric feature lifting to represent history-dependent residuals in linear andangular accelerations. Under the stated nonnegative-weight and structural constraints, each network output is convex with respect to the complete normalized and lifted causal feature history,which provides a structured residual model for online optimization. Furthermore, to reduce repeated neural-network evaluations over the prediction horizon, the history-conditioned networkoutputs and their current-node first-order sensitivities along the warm-start prediction trajectoryare computed in parallel on a GPU at each sampling instant. The resulting local affine models arethen passed to the RTI-SQP solver, thereby avoiding repeated evaluation of the complete temporalnetwork during the feedback phase and effectively decoupling neural-model computation from online optimization. Simulations with aerodynamic model mismatch and external-force disturbancesshow that, in the tested high-speed scenarios, IC-TCN achieves lower tracking errors than a standard TCN. In addition, on the reported desktop CPUGPU platform, the IC-TCN reduces thedata-driven preparation latency by 39.88% relative to the larger standard TCN.