X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model

Benchmark Model Rank Results
robot-manipulation-on-simplerenv-widow-xX-VLA-0.9B#1Average: 0.958
robot-manipulation-on-simplerenv-google-robotX-VLA-0.9B (Google Robot Visual Matching)#4Success Rate: 80.4
robot-manipulation-on-libero-longX-VLA-0.9B#14Success Rate: 97.6
robot-manipulation-on-robotwin-2-0-hard-50-tasksX-VLA-0.9B (RoboTwin-2.0 Hard)#33Success Rate: 39.0
robot-manipulation-on-robotwin-2-0-easy-50-tasksX-VLA-0.9B (RoboTwin-2.0 Easy)#35Success Rate: 70.0
robot-manipulation-on-libero-goalX-VLA-0.9B#39Success Rate: 97.8
robot-manipulation-on-liberoX-VLA-0.9B#47Success Rate: 98.1
robot-manipulation-on-libero-spatialX-VLA-0.9B#49Success Rate: 98.2
robot-manipulation-on-libero-objectX-VLA-0.9B#68Success Rate: 98.6
robot-manipulation-on-calvinX-VLA–avg. sequence length (D to D): 4.43