Alibaba's AMAP Lab publishes robot OS with lifelong graph memory

The framing on this one is more interesting than any single benchmark number. Researchers at Alibaba's AMAP CV Lab, the computer-vision team inside the company's mapping subsidiary, have posted ABot-AgentOS on arxiv, and the pitch is that long-horizon robot agents need something above their VLA controllers, not just better ones. They call that something an operating system.

What that means in practice, according to the paper, is scene-conditioned planning, context-isolated skill execution, multi-stage verification, and edge-cloud collaboration sitting on top of whatever perception-and-action model handles the low-level work. The piece that stands out is the memory design. Their Universal Multi-modal Graph Memory converts dialogue, visual observations, spatial context, temporal relations, and task traces into typed nodes and edges — one graph an agent can query across sessions rather than the usual per-episode context window.

To score the system the same group built EmbodiedWorldBench: 16 indoor, outdoor, and hybrid scenes, four difficulty levels, and over 200 tasks spanning navigation, object search, NPC dialogue, and dynamic events. On the memory-only side the paper reports 87.5 on LoCoMo and 88.6 on Mem-Gallery for the static configuration, ticking up to 88.7 and 89.0 once a self-evolution loop kicks in — a loop the paper says is engineered to avoid ground-truth leakage.

The honest caveat is that those numbers are self-reported on a benchmark the same authors defined and released in the same paper, which is normal for a foundation-tier drop but means the interesting comparisons — against something like RoboMemory or other lifelong-memory frameworks — are not in the paper, and neither is a deployment story on a named physical robot. AMAP has been open-sourcing an 'ABot' line for months (an ABot-M0 base model, an ABot-N0 navigation model), so treat this as the runtime piece slotting in beside the VLA pieces, not a standalone product.

Where it points is at a real strategic bet: if the durable layer in embodied AI turns out to be the OS above the model rather than the model itself, controlling that layer — with an open-source foothold and a benchmark you authored — is worth more than any single VLA leaderboard win.