Variational Cognitive Memory Abstraction for Long-Horizon Vision-Language-Action Models

Benchmark Model Rank Results
robot-manipulation-on-calvinCogMemory (ABC→D, Florence-2-L)–avg. sequence length (D to D): 4.64
robot-manipulation-on-calvinCogMemory (ABC→D, PaliGemma)–avg. sequence length (D to D): 2.92
robot-manipulation-on-calvinCogMemory (D→D, Florence-2-L)–avg. sequence length (D to D): 4.42
robot-manipulation-on-calvinCogMemory (D→D, PaliGemma)–avg. sequence length (D to D): 3.27
robot-manipulation-on-libero-goalCogMemory–Success Rate: 97.2
robot-manipulation-on-libero-longCogMemory–Success Rate: 97.1
robot-manipulation-on-libero-objectCogMemory–Success Rate: 98.8
robot-manipulation-on-libero-spatialCogMemory–Success Rate: 98.2