Maum AI's PonderPounce-WM hits 67.42% on RoboMME benchmark
The new robot foundation model adds a world model that forecasts how its surroundings will change 0.5, 1 and 2 seconds after its actions, Maum AI said.
Maum AI said its robotics research team WoRV has developed a new robot foundation model, PonderPounce-WM, which scored a 67.42% success rate on the global robot benchmark RoboMME in the company's own evaluation — a new record for the company.
The model pairs the earlier PonderPounce foundation model with world-model technology that predicts future situations. According to Maum AI, it forecasts how the surrounding environment will change 0.5, 1 and 2 seconds after its own actions.
A lightweight world-model structure predicts only the key visual features needed for decisions rather than generating full future video, which cuts computation. In a real-time asynchronous control evaluation using actual robot hardware, the model recorded a 67.44% success rate.
Targets and next steps
Maum AI aims to improve the model to reach the human-level RoboMME success rate of 90.50% by the end of the year. The company is also pursuing demonstrations in manufacturing, logistics, agriculture, construction and defense, with concrete results to be disclosed around year-end.
The world-model development was supported by South Korea's Ministry of Science and ICT under its "Physical AI Leading Technology Development" project.
In September, the previous PonderPounce model scored 60.83% on RoboMME, ranking first ahead of global robot AI models including US-based Physical Intelligence's Pi 0.5 and a Qwen-VL-based model from China's Alibaba.
Quick answers
What is PonderPounce-WM?
It is a robot foundation model developed by Maum AI's robotics research team WoRV, combining the earlier PonderPounce model with world-model technology that predicts future situations.
What score did PonderPounce-WM get on RoboMME?
It recorded a 67.42% success rate in Maum AI's own evaluation on the RoboMME benchmark, a company record. A real-time asynchronous control test on actual robot hardware recorded 67.44%.
What is Maum AI's target for the model?
Maum AI aims to raise the model to the human-level RoboMME success rate of 90.50% by the end of the year.