Track synthesis, materials and catalysis as it is published.
Search papers, labs, and topics across Lattice.
This paper introduces Achievement-Grounded Memory (AGM), a novel framework that enhances frozen vision-language-action (VLA) policies by enabling closed-loop execution through a structured subgoal sequence and progress verification. By utilizing proprioceptive cues and a lightweight verification head, AGM ensures that task progress is accurately tracked, mitigating the risks of persistent errors from open-loop actions. The approach significantly improves performance on the RoboMME Counting benchmark and physical robots, demonstrating that effective embodied memory hinges on disciplined state updates rather than sheer memory capacity.
Transforming open-loop action into a reliable closed-loop system, AGM achieves superior task performance while keeping VLA policies frozen and lightweight.
Frozen vision-language-action (VLA) policies offer broad manipulation skills but execute open-loop action chunks without tracking task progress, so the agent cannot reliably decide whether to continue, retry, or terminate. External memory is a natural remedy, yet it can be harmful when attempted actions are treated as completed progress, turning local execution errors into persistent task-state errors. We propose Achievement-Grounded Memory (AGM), a lightweight closed-loop framework for frozen VLA policies that represents a task as a subgoal sequence with a progress pointer and advances this memory only after the current subgoal is verified by physical evidence. Proprioceptive interaction cues decide when to verify, while coherent point tracking and language-conditioned cross-view comparison, sourced from frozen foundation models through a single 2.43M-parameter verification head, decide what was achieved. AGM thereby converts open-loop execution into a closed loop of execution, verification, and progress, keeping the policy frozen without test-time large-model inference. On the RoboMME Counting benchmark, AGM reaches on PickXTimes and on BinFill, surpassing the strongest memory-augmented baseline by points on average, and the framework yields equally decisive gains on a physical robot. Reliable embodied memory thus depends more on disciplined state updates than on memory capacity.