HiF-VLA: Hindsight, Insight and Foresight through Motion Representation for Vision-Language-Action Models

Benchmark Model Rank Results
robot-manipulation-on-calvinHiF-VLA (CALVIN ABC→D)#4avg. sequence length (D to D): 4.35
robot-manipulation-on-libero-spatialHiF-VLA#27Success Rate: 98.8
robot-manipulation-on-libero-longHiF-VLA#30Success Rate: 96.4
robot-manipulation-on-libero-objectHiF-VLA#42Success Rate: 99.4
robot-manipulation-on-libero-goalHiF-VLA#45Success Rate: 97.4
robot-manipulation-on-liberoHiF-VLA#51Success Rate: 98.0