MOST: Memory of Object States and Task Progress for Long-Horizon Mobile Manipulation
MOST: Memory of Object States and Task Progress for Long-Horizon Mobile Manipulation
This research presents MOST, a framework designed to enhance long-horizon mobile manipulation by compressing multi-view observations into temporal tokens that explicitly track object states and task progress. The authors introduce the RoboCasa-Long benchmark for multi-minute, multi-stage tasks. Experiments on RoboCasa365, RoboCasa-Long, and real-world robots show MOST outperforming the previous state-of-the-art model (Xiaomi-Robotics-1), raising average success rates from 65.9% to 72.1% on RoboCasa365, maximum task progress from 20.83% to 42.22% on RoboCasa-Long, and real-world success from 6.66% to 43.33%.
The paper introduces MOST, a policy-internal memory framework that compresses multi-view observations into temporal tokens supervised by object states and subtask progress, evaluated on the new RoboCasa-Long benchmark alongside RoboCasa365 and real-world experiments.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ti-stage mobile manipulation tasks. Across RoboCasa365, RoboCasa-Long, and real-world robot experiments, MOST consistently outperforms Xiaomi-Robotics-1 (XR-1), the previous state-of-the-art model. It raises average success from 65.9% to 72.1% on RoboCasa365, average maximum task progress from 20.83% to 42.22% on RoboCasa-Long, and real-world success from 6.66% to 43.33%. MOST also outperforms a matched visual-histor
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org