Search papers, labs, and topics across Lattice.
This paper introduces AURA-Mem, a novel memory architecture designed for robotic policies that optimizes memory usage by employing a learned gate to selectively write observations based on their impact on future actions. Unlike traditional KV-caches used in datacenters, which are inefficient for the continuous nature of robotic tasks, AURA-Mem maintains a constant memory size while significantly reducing write operations鈥攗p to 9.19 times fewer in easier scenarios鈥攚ithout sacrificing performance. The results demonstrate that AURA-Mem can match or exceed the performance of ungated policies while operating under strict memory constraints, making it a promising solution for edge robotics.
AURA-Mem achieves superior memory efficiency for robotic policies by intelligently controlling write operations, outperforming traditional KV-caches in both accuracy and resource usage.
The KV-cache is the right memory for datacenters but the wrong memory for robots. Datacenter inference batches many short requests and resets them, amortizing an attention cache across a crowd. Embodied agents instead run one long, non-resetting episode on bandwidth-limited edge hardware, where high-bandwidth memory and flash are scarce, flash has finite write endurance, and memory writes rather than compute can become the binding constraint. AURA-Mem (Action-Utility Recurrent Adaptive Memory) targets this regime. It wraps a frozen vision-language-action backbone with a constant-size recurrent memory and a learned gate that writes only when the current observation would change the next action: memory that knows when to stay silent. Unlike reconstruction-based memory, the gate is trained directly against a closed-loop action-error signal. Its inference state is fixed at 4,224 bytes regardless of horizon, while a KV-cache grows to 6,061 times larger at 100,000 steps. On a controlled synthetic benchmark, AURA-Mem matches the best O(1) baseline in accuracy while using 5.19-6.13 times fewer writes, and up to 9.19 times fewer writes on easier configurations. Budget-matched random and periodic schedules do not recover this gain, isolating the benefit to the action-surprise signal. On a trained closed-loop OpenVLA-OFT 7B panel on LIBERO-Long (n=60 episodes per arm), the gate does not hurt success: AURA-Mem matches the ungated base policy (0.233) and slightly exceeds an always-write KV arm (0.217), while using 7.0 times fewer writes and constant memory. We also instantiate an approximate-information-state value-loss bound as a methodology demonstration; at this scale, the bound is vacuous rather than a guarantee.
