AGM: Achievement-Grounded Memory for Closed-Loop Agents with Frozen VLA Policies
AGM only records progress after a robot verifies that the subgoal actually happened.
The paper targets frozen VLA policies that run action chunks open-loop and can mistake attempted actions for completed work. AGM adds a subgoal pointer, interaction cues for when to check, and a small verification head using frozen foundation-model signals to decide what was achieved. The authors say this turns execution into an execute-verify-progress loop without updating the policy or using large-model inference at test time. They report gains over memory-augmented baselines on RoboMME Counting and on a physical robot. ArXiv · AI/CL/LG's note
The paper targets frozen VLA policies that run action chunks open-loop and can mistake attempted actions for completed work. AGM adds a subgoal pointer, interaction cues for when to check, and a small verification head using frozen foundation-model signals to decide what was achieved. The authors say this turns execution into an execute-verify-progress loop without updating the policy or using large-model inference at test time. They report gains over memory-augmented baselines on RoboMME Counting and on a physical robot. ArXiv · AI/CL/LG's note
score 5