Social behavioral alignment
Built from a counterfactual benchmark we construct.
The robot starts at the crosswalk and heads in. The base model always goes around — but it gives a person the same small clearance it would give a box. Cycle through three different-looking pedestrians: SALSA opens a wider berth for every one of them — it responds to person-ness, not appearance. Swap in an object and SALSA stays just a hair beyond base.
Implementation: counterfactual person/object pairs — same scene, same spot, identity flipped — are the only supervision. The policy can lower its loss only by responding to person-ness, so social clearance emerges with no human labels.