MemBodied: Recurrent Associative Memory for Vision-Language-Action Models
MemBodied adds fixed-size episodic memory so robot policies can use earlier scene information without carrying full observation history.
The paper targets manipulation tasks where the needed clue may appear only in a past observation. Its memory has an associative state across policy calls and an episode anchor for the initial scene. In five RMBench memory tasks, it reports 7.81x the mean success rate of a stateless policy and 2.98x vanilla recurrent memory. On LIBERO-Long, it reached 90.6%, 5.4 points above the stateless pi_0 policy. ArXiv · AI/CL/LG's note
The paper targets manipulation tasks where the needed clue may appear only in a past observation. Its memory has an associative state across policy calls and an episode anchor for the initial scene. In five RMBench memory tasks, it reports 7.81x the mean success rate of a stateless policy and 2.98x vanilla recurrent memory. On LIBERO-Long, it reached 90.6%, 5.4 points above the stateless pi_0 policy. ArXiv · AI/CL/LG's note
score 5