Maum AI's PonderPounce-WM Robot Model Hits 67.42% on RoboMME

South Korean company Maum AI said on November 1 that its new robot foundation model, PonderPounce-WM, reached a 67.42% success rate in its own evaluation based on the global RoboMME benchmark, a new internal record. The model was developed by the company's robotics research team, WoRV.

PonderPounce-WM combines the earlier PonderPounce model with a world model that predicts future situations. In September, PonderPounce scored 60.83% on RoboMME and, according to the company, ranked first ahead of models such as Physical Intelligence's Pi 0.5 and an Alibaba Qwen-VL-based model. The new version improves on that by 6.59 percentage points within a few weeks.

Predicting what happens next

The earlier model chose its next action from the current scene and past task information. The new one also predicts how the surroundings will look 0.5, 1 and 2 seconds after its own action. Maum AI aimed this at complex tasks that require remembering context, such as picking up an object whose position it noted before it was hidden, or moving a set number of items and then stopping.

To keep the extra computation from slowing the robot, the company did not generate full video of future scenes. Its lightweight world model predicts only the key visual features needed for decisions.

On real robot hardware, in a real-time asynchronous control test where perception, judgment and action run at the same time while the robot moves, the model recorded a 67.44% success rate.

Targets for the year-end

Maum AI plans to keep refining the model until it reaches the human benchmark of 90.50% on RoboMME by the end of the year. It is also pursuing field trials in manufacturing, logistics, agriculture, construction and defense, and expects to publish concrete results around year-end.

The development was supported by the Ministry of Science and ICT's Physical AI leading technology development program. CEO Choi Hong-sub, who leads technology, said adding a world model let the robot imagine the future and improved performance, and that the company will keep improving toward human-level results on the benchmark.