A robotics team has taught humanoid robots to gesture around obstacles instead of just bumping into them mid-sentence.
Researchers built GestAdapt, a framework that generates co-speech gestures for robots while factoring in the physical workspace around them, such as a nearby wall. It trained on six co-speech gesture datasets through a shared motion representation, then retargeted the output to different robot bodies, including the Reachy2 humanoid. In a user study, workspace-aware gestures scored 3.24 out of 5 for quality, versus 2.43 for a version that ignored workspace limits, though both trailed the 3.68 score for real human reference motions. On an actual Reachy2 robot, the workspace-aware gestures beat both the no-workspace baseline and ground-truth motions retargeted after the fact, winning 69.7% of head-to-head comparisons.
The result reframes a common robotics shortcut: generate a gesture, then clip or warp it to fit the room. This paper shows planning around constraints from the start beats fixing them after, which matters as humanoid robots move from lab demos into offices, warehouses, and homes full of walls, desks, and people standing too close.
It is still a modest gain over no constraints at all, and nowhere near matching a human talking with their hands - a reminder that gesture generation remains harder than it looks in a conference demo.