Long-Horizon System Identification of Industrial Manipulators: A Comparative Study Using Measured-Input Rollouts

Accurate dynamic models are essential for high-performance tracking, robust feedback, and model-based planning on industrial manipulators. In practice, however, identification pipelines are often tuned for low one-step prediction error, even though models are deployed by iterating the dynamics, causing errors to accumulate over time. This work studies plant-level, torque-driven system identification on two 7-DoF arms (KUKA LWR iiwa and Baxter) using real joint-state and torque logs. Next-step learning is posed as xt+1=fθ xt, ut) with historywindowed regressors, and models are evaluated using both single-step metrics and measured-input long-horizon rollouts. We compare a linear statespace surrogate, a history-linear ARX model, and compact NARX-MLP and NARX-LSTM models under trajectory-level in-/out-of-distribution splits with multi-seed confidence intervals. While ARX achieves the lowest one-step RMSE, the NARXMLP exhibits markedly slower error growth over rollouts up to 1000 steps and residuals closest to white noise, making it more suitable for iterative use in simulation and MPC. These trends persist under ID → OOD shift, highlighting the importance of measured-input long-horizon evaluation.