Accurate learned dynamics are essential for tracking, feedback, and planning on industrial manipulators, yet single–step prediction error often fails to reflect how models behave when iterated in simulation or control. We study plant–level, torque–driven system identification on a 7-DoF KUKA LWR iiwa using real joint state and torque logs. Two linear baselines are compared: a memoryless linear state–space surrogate (LS–SS) and a history–linear Auto-regressive with Exogenous Input (ARX) model trained with ridge regression. Beyond standard next–step metrics, we evaluate models with measured–input long-horizon rollouts, trajectory-level in/out-of-distribution (ID/OOD) splits, and 95% confidence intervals across seeds. On KUKA, ARX consistently achieves the lowest one–step RMSE and exhibits slower error growth over horizons up to 1000 steps than LS–SS, although both accumulate error with horizon. These results hold under an ID→OOD shift constructed from the most aggressive trajectory in the set. Our contributions are: (i) a reproducible plant–level system identification (SysID) protocol for manipulators that complements single–step metrics with measured–input rollouts and residual tests; and (ii) a head-to-head analysis showing that short history in a linear ARX estimator materially improves iterative stability relative to an H=1 surrogate. We conclude that one–step accuracy alone is insufficient for models intended for iterative use and recommend long-horizon, measured-input evaluation as standard practice for manipulator SysID.
Haque et al. (Thu,) studied this question.