SimpleMemVLA: זיכרון פשוט ומוצלח למודלי תצלום-לשון-פעולה

תקציר מקורי באנגליתarXiv:2609.05533v2 Announce Type: replace-cross Abstract: Long-horizon manipulation is partially observable: the information needed to choose the next action may appear only in observations from minutes earlier. Existing memory mechanisms for VLAs, such as retrieval banks, learned compressors and recurrent states, must decide what to keep from the past before knowing what a future decision will require. They were motivated by the assumption that minute-scale history is too large to process directly, which no longer holds for modern VLM backbones. We propose SimpleMemVLA, a VLA without a dedicated memory module that uses the backbone's native video context directly as memory. It keeps the sampled history intact in the timestamped video format the backbone was pretrained to process, routes t