StableVLA: Towards Robust Vision-Language-Action Models without Extra Data

Benchmark Model Rank Results
robot-manipulation-on-libero-goalStableVLA (0.5B, clean)–Success Rate: 98.0
robot-manipulation-on-libero-longStableVLA (0.5B, clean)–Success Rate: 93.6
robot-manipulation-on-libero-objectStableVLA (0.5B, clean)–Success Rate: 98.8
robot-manipulation-on-libero-spatialStableVLA (0.5B, clean)–Success Rate: 96.2