R2-WAM: What Visual Representations Do World Action Models Need?

Loading video
Loading videoA controlled study of visual representations, five integration mechanisms and feature combinations for World Action Models: general-purpose representations give the broadest robustness, channel-wise joint prediction scores highest, and R2-WAM reaches 84.7% on RoboCasa GR1 Tabletop and 92.8% on RoboTwin 2.0.
Category: humanoid
Author: @ai_bites
Date: 2026-10-15T00:00:00
Duration: 93.766s
Reference: https://arxiv.org/abs/2610.13024





