Robots can pick up the cup. They cannot share the room.

The cooperation gap in foundation robot models is the next deployment bottleneck—a framing that reflects one interpretive projection of the gap’s significance, not independently confirmed by peer review or industry consensus—and the field has been grading the wrong problem. Dexterity and social cooperation are not the same skill. A policy that picks up a cup is not on the scoreboard for a hand-off, a turn, or a person stepping into its workspace. Industry has been shipping one while the operators are asking for the other.

The HRIBench authors' adaptation study gives that gap its first number. GR00T N1.5 posts a 0.10 success rate on physical tasks that require sharing a workspace with a person, climbing to 0.43 after retraining on the benchmark's simulation data. HRIBench formalizes three roles, the human partner, the teammate, and the disruptor, and runs more than 650 episodes against adapted versions of GR00T, pi0.5, and ACT. The lift is real and small, the kind of number that says the curriculum did not exist, not that the students are slow.

The mechanism is portable. Every shared-workspace deployment will hit the same wall: a model that can grasp a bolt and miss the turn. HRIBench turns the vague claim "robots struggle with people" into a measurable gap with a baseline and a lift.

Who loses if the gap stays open: operators, safety leads, and the buyers told a humanoid is one fine-tune from the loading dock. Who gains when it closes: the lab that ships the first policy whose cooperation score is part of the spec sheet, not an afterthought.