BayesianVLA: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries

Benchmark Model Rank Results
robot-manipulation-on-robocasa-gr1-tabletopLangForce + Qwen3-VL-4B#6Success Rate: 52.6
robot-manipulation-on-libero-goalLangForce (LIBERO-Goal)#7Success Rate: 99.4
robot-manipulation-on-libero-spatialLangForce (LIBERO-Spatial)#14Success Rate: 99.2
robot-manipulation-on-simplerenv-widow-xLangForce + Qwen3-VL-4B#23Average: 0.665
robot-manipulation-on-libero-objectLangForce (LIBERO-Object)#30Success Rate: 99.6
robot-manipulation-on-liberoLangForce#35Success Rate: 98.4
robot-manipulation-on-libero-longLangForce (LIBERO-Long)#46Success Rate: 95.2