Fig. 6
Distributional comparison of test-set metrics for inverse kinematics and dynamics across all evaluated methods. The figure illustrates model performance in terms of accuracy, computational cost, and uncertainty calibration. (a) Joint-angle Root Mean Square Error (RMSE) in radians. (b) End-effector position error in meters. (c) Per-step inference latency in milliseconds. (d) Negative log-likelihood of task-space measurements in nats. (e) Empirical coverage probability. For each plot, the colored curves represent the kernel density estimates for a specific model. The shaded background shows the empirical reference distribution across the entire test dataset, providing a baseline for comparison. The dotted vertical lines indicate a target value; for latency (c), this line represents a critical real-time performance threshold required by the control loop. The concentrated and favorably centered distributions for our ILNN model demonstrate its superior accuracy, computational efficiency, and better-calibrated uncertainty quantification compared to the baseline approaches.
