# Shanghai Jiao Tong University and others propose VaM: using value prediction for task-progress memory, achieving over 95% success rate in 4-hour continuous carton assembly

- Source: 机器之心 · 微信公众号
- Published: 2026-10-30T04:22:49.000Z
- RoboHot: https://robohot.news/en/items/dq8y756hpsy7fvfjxyap17nlw
- Original: https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2651071094&idx=2&sn=b32d665f396ca7ea913fa843efb9938c&chksm=8523b31086fb7b5ae580eacbb288c8996b2adfb7f593890451d6912b8586cc47fb1b2fbcfbee&scene=126&sessionid=1793371799#rd

## Summary

Shanghai Jiao Tong University, Shanghai Innovation Institute, and others propose Value as Memory (VaM), which feeds the recent prediction sequence of a value model into the action policy as task-progress memory, combined with an offline-to-online real-robot reinforcement learning pipeline. In a carton assembly task, VaM ran autonomously and continuously for 4 hours with a task success rate above 95%. Ablations show that compared with a version without the value cache, the final task success rate improved by 29.4%; the online RL approach improved average success rate by 32.8% across three tasks and reduced the real-world time cost per gradient step by 57.8% on average.