SimpleMemVLA Datasets

This repository contains the LeRobot v3 datasets used to train and evaluate SimpleMemVLA, a vision-language-action model with native-video memory for long-horizon robotic manipulation. The datasets cover five benchmarks:

• RMBench (yinchenghust/rmbench_lerobot)
• RoboMME (yinchenghust/robomme_lerobot)
• MIKASA-Robo (yinchenghust/mikasa_lerobot)
• RoboMemArena (yinchenghust/robomemarena_lerobot)
• LIBERO (yinchenghust/libero_lerobot)

Each dataset includes per-frame subtask supervision and is provided in LeRobot format. For details on the method, see the paper. Code is available at github.com/wadeKeith/SimpleMemVLA. Project page: https://huggingface.co/collections/yinchenghust/simplememvla