MemBodied: Recurrent Associative Memory for Vision-Language-Action Models
MemBodied gives robot policies a fixed-size episodic memory instead of feeding them an expanding history.
The paper targets manipulation tasks where the robot needs information seen earlier in the episode but not present in the current view. Its memory has an associative state carried across policy calls and an anchor for the initial scene. On five RMBench memory tasks, it reports 7.81x the mean success rate of a stateless policy and 2.98x vanilla recurrent memory. It also reports 90.6% on LIBERO-Long, 5.4 points above the stateless pi_0 policy. Source: HF Daily Papers' note.
The paper targets manipulation tasks where the robot needs information seen earlier in the episode but not present in the current view. Its memory has an associative state carried across policy calls and an anchor for the initial scene. On five RMBench memory tasks, it reports 7.81x the mean success rate of a stateless policy and 2.98x vanilla recurrent memory. It also reports 90.6% on LIBERO-Long, 5.4 points above the stateless pi_0 policy. Source: HF Daily Papers' note.
score 4