Adding Motion Sense to VLAs With Frozen SAM 2 Lets Robots Catch Moving Bottles

nikhilaravi · x · 2026-10-08

Position alone isn't enough for dynamic robotic tasks. By using a frozen SAM 2 video model to distill recent frames into a compact motion cue (speed and direction) for a VLA, researchers enabled the robot to actually catch a moving bottle. SAM 2 is open-sourced.

Original post →

More from Embodied

Embodied channel →