AURA: Action-Gated Memory for Robot Policies at Constant VRAM
The paper presents AURA-Mem, a novel memory architecture designed for robotic policies that operates with constant VRAM. This approach utilizes an action-gated mechanism to optimize memory writes, significantly reducing the number of writes needed compared to traditional methods. The results indicate that AURA-Mem maintains accuracy while using fewer resources, making it suitable for bandwidth-limited edge hardware.
- ▪AURA-Mem targets the limitations of memory in robots, contrasting with datacenter memory systems.
- ▪The architecture uses a learned gate that writes to memory only when necessary, reducing unnecessary writes.
- ▪In benchmarks, AURA-Mem achieves comparable accuracy to existing methods while using significantly fewer memory writes.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2606.02775 (cs) [Submitted on 1 Jun 2026] Title:AURA: Action-Gated Memory for Robot Policies at Constant VRAM Authors:Josef Chen View a PDF of the paper titled AURA: Action-Gated Memory for Robot Policies at Constant VRAM, by Josef Chen View PDF HTML (experimental) Abstract: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.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
