Safe Social Navigation
Act on What You See
A VLA navigator already tells social entities from obstacles and sees danger coming — it just doesn't act on either. Follow the dot from the problem, through SALSA's two alignment stages, to the result.
01
The problem
VLA navigators do avoid — by size,
not meaning, and far too late
02
The diagnosis
they already perceive them —
it never reaches the action
03
Social alignment
counterfactuals teach
a person ≠ an object
04
Temporal safety
hindsight builds a three-tier
normal → slow → stop
RESULT
-86%
near-collisions
clearance +75.9% · social 53→93%
avoids — but late, and by size
not meaning · a person = a cone
probe by layer · peak ~L9
person ≠ object
danger ahead
action
decoded inside · never acted on
wide for a person, tight for a cone
CF-Acc 53% → 93%
slow
stop
hindsight → a three-tier response
90% of danger frames resolved
-86%
near-collisions
clearance +75.9% · social 53→93%
❙❙ Pause