Top AI Product

Teaching a humanoid to pick up a box and carry it across a room usually means hours of teleoperation or motion capture. NVIDIA’s GRAIL goes the other way: it generates loco-manipulation behavior — locomotion plus object handling — directly from 3D assets and motion priors pulled from video.

## The data bottleneck it targets

Humanoid robots are starved for training data. Every new object or task tends to need fresh human demonstrations, and that doesn’t scale. GRAIL’s bet is that you can synthesize the behavior instead: combine a 3D model of the object with priors learned from how people move in video, and produce usable whole-body motion without a human in the loop for each new case.

## Why it matters

This is the same thread NVIDIA keeps pulling across physical AI — world models, simulation, and now skill generation. If loco-manipulation data can be generated rather than captured, the slowest, most expensive part of training humanoids gets dramatically faster.