Dual Latent Memory in Vision-Language-Action Models for Robotic Manipulation

The paper introduces LaMem-VLA, a latent-memory-native framework for long-horizon robotic manipulation. Unlike approaches that append observation windows or retrieve history as auxiliary policy context, LaMem-VLA reconstructs retrieved experience into latent memory tokens within the VLA model’s native embedding space. A curator maintains short-term and long-term memory vaults, a seeker retrieves relevant evidence, a condenser compresses it into latent tokens, and a weaver interleaves those tokens with the current observation and instruction. The authors report experiments on SimplerEnv and LIBERO, but the provided abstract does not include quantitative results.