流式推理优化，让机器人决策延迟低至0.68秒

SimpleMemVLA, developed by 面壁智能 with six partner institutions, treats timestamped visual history as native context instead of a dedicated memory module, achieving SOTA across four memory benchmarks and enabling visual context learning without parameter updates. On RoboMemArena, its full-stage task success reaches 63.6% with a 126-second history window (up 17.4 points), boosting counting tasks from 31.4% to 71.4%, while streaming inference reuses key-value cache to cut decision latency for 60-second histories from 1.02s to 0.68s on a single H100. The system retains up to 262k tokens of context, uses plaintext timestamps to preserve frame order, and demonstrates in real-robot tests that it can locate occluded objects by recalling pre-occlusion visual history.

why now

机器人决策延迟降至0.68秒，开源VLA记忆模型发布。

topic

AI Tech & New Models

source

量子位（公众号）