Robots trained to grab objects from one camera angle often fall apart when you move the camera.
Researchers built a system called InfiNoVA that fixes this without collecting more physical footage. It takes synchronized multi-camera recordings of a robot performing a task, rebuilds each moment as a 3D scene using Gaussian splatting, and renders synthetic training views from camera angles that were never actually filmed, all while keeping the recorded actions correctly matched to what the robot sees. Across four real-world manipulation tasks, robots trained on this synthetic data succeeded 5.4 times more often from random, unseen viewpoints than robots trained without augmentation or with a prior method called VISTA. It even beat robots trained on footage from all five physical cameras, by a factor of 1.7.
The expensive part of robot learning has always been the data: every camera angle you want a robot to handle usually means physically re-recording demonstrations from that angle. If synthesizing consistent viewpoints works as well as this suggests, it could make camera placement a non-issue for deployment, without redesigning the policy itself, which is the kind of unglamorous infrastructure fix that actually gets robots out of controlled labs.
The gains are real, but they come from four tasks in one lab's setup. The distance between an arXiv benchmark and a warehouse robot that shrugs off a bumped camera is still considerable.