for each subtask (drop-off and pickup). The training results showed 0% FPs and roughly 5.2% FNs for the first NN and 0% FPs and approximately 1.2% FNs for the second one. ra = 0.3 m 1 1 0.5 0.5 0 -0.5 y (m) 1 0 -0.5 -1 -1 0 1 2 x (m) 3 4 0 -0.5 -1 0 1 2 x (m) 3 4 0 3 (c) ra = 0.45 m ra = 0.7 m ra = 0.7 m 1 1 0.5 0.5 0 -0.5 0 0.5 -1 2 x (m) (d) Safe Labels 3 0 0.5 4 Unsafe Labels y (m) 1 1 2 x (m) (b) 0.5 0 1 (a) y (m) y (m) ra = 0.45 m 0.5 y (m) y (m) ra = 0.3 m In Figure 4, the initial positions in the training set for the drop-off and pickup operations are presented with their respective labels and the NN results inside the proposed workspace. Each subfigure denotes a different avoidance 4 0 -0.5 -1 1 NN Safe Labels 1.5 x (m) (e) 2 0 2.5 NN Unsafe Labels 1 Drop-Off Position 2 x (m) (f) 3 4 Obstacles Positions Figure 4. Safety maps for the initial and final positions in training sets with different avoidance distances. The (a) drop-off task and (b) pickup mission for ra = 0.3 m. The (c) drop-off task and (d) pickup mission for ra = 0.45 m. The (e) drop-off task and (f) pickup mission for ra = 0.7 m. Verified (NN Output: Safe; Plant: Safe) Verified (NN Output: Unsafe) Unknown (NN Output: Safe; Plant: ?) Not Verified (NN Output: Safe; Plant: Unsafe) 0.5 Verisig Test Area 1 y (m) y (m) 0.5 0 0 -0.5 -1 0 1 2 x (m) (a) 3 -0.5 0 0.2 0.4 0.6 0.8 1 x (m) (b) Figure 5. (a) The pickup task with avoid distance ra = 0.45 m in Figure 4(c). (b) The initial set was divided into small subsets and verified. No unsafe sets were obtained. 108 * IEEE ROBOTICS & AUTOMATION MAGAZINE * JUNE 2020