MemMimic is a non-Markovian benchmark for robotic manipulation tasks introduced in the paper "Gated Memory Policy". The dataset features tasks with varying memory requirements, ranging from Markovian tasks to non-Markovian tasks that depend on historical information spanning single or multiple interaction trials.
The dataset is designed to evaluate visuomotor policies on their ability to selectively recall and process history context, particularly addressing issues like distribution shift and overfitting when extending observation histories.
MemMimic is a non-Markovian benchmark for robotic manipulation tasks introduced in the paper "Gated Memory Policy". The dataset features tasks with varying memory requirements, ranging from Markovian tasks to non-Markovian tasks that depend on historical information spanning single or multiple interaction trials.
The dataset is designed to evaluate visuomotor policies on their ability to selectively recall and process history context, particularly addressing issues like distribution shift and overfitting when extending observation histories.
