New Vision-Language Model Achieves 76.6% Success on R2R-CE Benchmark
July 8, 2026
This model achieves 76.6% success on the R2R-CE unseen validation set, outperforming the top single-camera approach by 9.7 percentage points. It demonstrates superior instruction-following capabilities in novel environments while utilizing reduced sensing inputs.
HOW THIS AFFECTS YOU
●
researcherThis establishes a new state-of-the-art for embodied instruction following in unseen environments.