Bird's-eye View
In this section we present a selection of videos from the agent using the bird's-eye view input. On videos our agent is marked green while pale blue is the original human drive (it is added only in for analysis and is inaccessible to the agent).
In this section we present a selection of videos from the agent using the bird's-eye view input. On videos our agent is marked green while pale blue is the original human drive (it is added only in for analysis and is inaccessible to the agent).
Below there is a semantic saliency map, the grayscale indicate how much the agent's policy is influenced by other traffic participants (the more white the more).
Below there is a semantic saliency map, the grayscale indicate how much the agent's policy is influenced by other traffic participants (the more white the more).
NGSIM - successful cases
NGSIM - successful cases
success, smooth lane change 28.07.2020_17:06:17_28_ckpt=11750_scenario=NGSIM_LANE_CHANGE_US101_VALIDATION_steps=27_total-reward=1.9.avi
Smoothly executed
Smoothly executed
(no slowdown at the end)
success, dynamic lane change 28.07.2020_17:06:17_19_ckpt=11750_scenario=NGSIM_LANE_CHANGE_I80_VALIDATION_steps=26_total-reward=1.9.avi
Slightly impatient but well done
Slightly impatient but well done
(looks a bit artificial, low speed)
NGSIM - failed cases
NGSIM - failed cases
failed, couldnt decide, to fast at the end 28.07.2020_17:06:17_17_ckpt=11750_scenario=NGSIM_LANE_CHANGE_US101_VALIDATION_steps=82_total-reward=-0.30000000000000004.avi
Decided to take try with another gap
Decided to take try with another gap
(seemingly was 'afraid' of collision in the first gap although could easily fit in, but in the second gap the agent expected timeout penalty, hence such speed-up)
failed, unnecesary slow down 28.07.2020_17:06:17_19_ckpt=11750_scenario=NGSIM_LANE_CHANGE_US101_VALIDATION_steps=38_total-reward=-0.19999999999999996.avi
Slowing down, sloppy maneuver
Slowing down, sloppy maneuver
(speed controlling issues)
openDD - successful cases
openDD - successful cases
success, reference waits, agent enters 28.07.2020_15:12:48_3_ckpt=05300_scenario=OPENDD_RDB7_DENSE_steps=71_total-reward=2.0.avi
Leaves reference driver behind
Leaves reference driver behind
(common case, also speed-up at exit)
success, almost like reference driver 28.07.2020_15:12:48_9_ckpt=05300_scenario=OPENDD_RDB7_DENSE_steps=136_total-reward=2.0000000000000004.avi
Almost identical to reference driver
Almost identical to reference driver
(speed-up at exit)
openDD - failed cases
openDD - failed cases
failed, enforcement of priority 28.07.2020_15:12:48_8_ckpt=05300_scenario=OPENDD_RDB4_DENSE_steps=32_total-reward=-0.8.avi
Enforcement of priority
Enforcement of priority
(could end up well if only acceleration would be stronger)
failed, stuck between two 28.07.2020_15:12:48_5_ckpt=05300_scenario=OPENDD_RDB4_DENSE_steps=74_total-reward=-0.5.avi
Collision, human driver enforces priority
Collision, human driver enforces priority
(reminder: human drivers are not interactive, agent shouldn't be there, so if he'd behave like the reference driver, he wouldn't face such an inevitable situation - in between two human drivers