PonderPounce: A Pretrained MLLM as an Episode Context Engine for Robot Control
PonderPounce uses an MLLM’s own context window as robot episode memory, instead of adding a separate memory module.
The system pairs Ponder, a slower MLLM that accumulates observations and demonstrations, with Pounce, a faster VLA action model. Pounce receives only the newest continuous cognition token and its age from Ponder while acting from current observation, instruction, and proprioception. The authors report 20Hz action playback, with 78ms median cognition refresh and 25ms action invocation. On RoboMME, the 9B version scores 60.83% with base-scale data and 75.54% with 9x data. Source: HF Daily Papers' note.
The system pairs Ponder, a slower MLLM that accumulates observations and demonstrations, with Pounce, a faster VLA action model. Pounce receives only the newest continuous cognition token and its age from Ponder while acting from current observation, instruction, and proprioception. The authors report 20Hz action playback, with 78ms median cognition refresh and 25ms action invocation. On RoboMME, the 9B version scores 60.83% with base-scale data and 75.54% with 9x data. Source: HF Daily Papers' note.
score 4