Web Services (https://aws.amazon.com). Each subset took roughly an hour to verify, although some took longer due to branching introduced by if-cases in the planner. Figure 5 closely matches the corresponding graph in Figure 4(c). We confirmed that the NN was, indeed, conservative Actual Path Desired Trajectory Waypoints Goal Obstacles 2 2 1 y (m) 1 y (m) so that no unsafe events occurred when its output was safe. The same procedure can be used for all other cases. Note that we verified the safety for only the first NN as a proof of concept; the procedure for the second NN would be exactly the same. 0 0 -1 -1 -2 -2 -4 -2 0 x (m) 2 4 -4 -2 ra = 0.4: Unsafe ra = 0.5: Unsafe ra = 0.6: Safe Wind 2 1 1 y (m) y (m) 4 (b) 2 0 0 -1 -1 -2 -2 -2 2 ra = 0.4: Unsafe ra = 0.5: Unsafe ra = 0.6: Unsafe ra = 0.7: Safe Wind (a) -4 0 x (m) 0 x (m) 2 Wind 4 ra = 0.4: Unsafe ra = 0.5: Safe (c) -4 -2 0 x (m) Wind 2 4 ra = 0.4: Unsafe ra = 0.5: Unsafe ra = 0.6: Safe (d) Figure 7. The experimental results in which the trained NN was used to make safety decisions and replan accordingly. The desired versus the actual trajectories are presented in the top row of subfigures, with the relative snapshots from the experiments in the bottom row. (a) The round 1 pickup task. (b) The round 1 drop-off task. (c) The round 2 pickup task. (d) The round 2 drop-off task. 110 * IEEE ROBOTICS & AUTOMATION MAGAZINE * JUNE 2020https://aws.amazon.com