Maum AI world-model upgrade lifts its RoboMME robot memory score to 67.4%

Seoul-based Maum AI said on Oct 29 that PonderPounce-WM, a version of its robot foundation model with an added world model, reached a 67.42% success rate on RoboMME, a memory benchmark for robot manipulation built by the University of Michigan, Stanford University and Figure AI, and 67.44% when run asynchronously on real robot hardware, TechWorld reported.

Maum AI has not posted the announcement on its own site and the new score is not yet on the RoboMME leaderboard, so both figures are unconfirmed. The earlier PonderPounce scored 60.83%. The world model predicts visual features and state changes 0.5, 1 and 2 seconds ahead instead of generating full future video, which the company says keeps real-time control possible; the work was done under the science ministry's physical-AI leading-technology program, four months after the award.

MEKAWIRE check of the RoboMME leaderboard (data updated Oct 21): PonderPounce's 60.83% sits eighth, below the human reference, an oracle baseline and five AI systems, among them the open-source SimpleMemVLA at 88.25% and the proprietary 3-tier Astra at 79.12%.

A 67.42% score would therefore not top the open-model ranking that Maum AI said it led on Sept 29. Maum AI also showed a quadruped landmine detector this month.