WorldBridge is a robot model that pairs video prediction with action learning to handle tasks it wasn't trained on.
Key finding: On RoboTwin 2.0 out-of-distribution tasks, success rose from 25.07% with in-distribution video-action training to 55.90% with extra videos, and to 70.70% with one action-labeled demo per task.
Architecture: Separate video and action experts linked by structured attention, trained with joint flow matching, with visual memory at two temporal scales and action forcing.
Benchmarks: The authors report state-of-the-art among compared methods: 96.23% on RoboTwin 2.0 full data, 80.9% on clean to random, 82.8% on RoboCasa GR1 TableTop, and 57.6% on DOMINO Level-1.
Sign in to comment.
TL;DR of the paper, generated automatically.
WorldBridge is a robot model that pairs video prediction with action learning to handle tasks it wasn't trained on.