Splat2Act Customizes a Robot to an Interior Based on a Scan on the Phone

A digital twin produces the demonstrations needed for adaptation. In the trial on a Stretch 3 robot, the personalized policy succeeds in 63.9% of the shots.

Splat2Act transforms an indoor scan and images of objects into a simulated environment, then generates take demonstrations to adapt a vision-language-action policy. The pipeline spans 101 campus spaces and produces more than 68,000 demonstrations.

In the captured deployment scenes, the initial generalist policy fails at the evaluated holds before adaptation. The best customized version achieves 62.5% in simulation and 63.9% on twelve tasks performed with a Stretch 3. The result illustrates a targeted adaptation to given places and objects, not universal home autonomy.

Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.