Simulaton Results: SIGMO and Naive MPPI on the identical seeds in two interaction-intensive scenarios: a merge into traffic and a narrow passage shared with pedestrians.
Naive MPPI: Stays in the source lane
Remains in its original lane, unable to initiate a merge without an available gap.
SIGMO: Nudges into the gap
Ego gradually approaches the target lane, adjacent vehicle slows and its estimated yield probability rises as it slows, and the ego merges ahead.
Naive MPPI: Stranded in the corridor
Reacts only to imminent collisions; brakes repeatedly for crossing pedestrians, waits over 10 s as pedestrians cross
SIGMO: Passes, then yields
Adjusts its speed to progress through pedestrian traffic that are likely to yield while slowing for an oncoming pedestrian.
Hardware Results: SIGMO on a 1/10-scale F1Tenth car merging from a ramp while a pedestrian walks the lane, running at alongside onboard perception and localization. One Naive MPPI run for comparison, then three SIGMO runs showing it yielding after a short or a long attempt or merging ahead proactively.
Naive MPPI stalls at the merge.
No interaction term; waits on the ramp for the pedestrian to pass.
SIGMO Surrender Short:
Tests the gap, sees no yield, gives way within a second or two.
SIGMO Surrender Long:
Presses the merge for several seconds, the pedestrian holds course, SIGMO backs off and merges behind
SIGMO Proactive:
Nudges and merges ahead
Reads the pedestrian as likely to yield, claims the gap, merges in front.