SAM3D-Guided Object-Centric Representation Alignment for Vision-Language-Action Models

Benchmark Model Rank Results
robot-manipulation-on-calvinSAM3D-VLA–avg. sequence length (D to D): 4.11
robot-manipulation-on-liberoSAM3D-VLA–Success Rate: 99.1
robot-manipulation-on-libero-goalSAM3D-VLA–Success Rate: 99.1
robot-manipulation-on-libero-longSAM3D-VLA–Success Rate: 98.4
robot-manipulation-on-libero-objectSAM3D-VLA–Success Rate: 99.7
robot-manipulation-on-libero-spatialSAM3D-VLA–Success Rate: 99.2