arXiv Robotics — research abstracts· Hao Wu, Tianyu Shao, Shengju Qian, Yizhou Zhao, Fan Zhang, Penghao Zhao, Weiyan Wang, Yufei Huang, Fan Xu, Kun Wang, Xin Wang, Ming-Hsuan Yang, Yuxuan Liang·· 8 hours agoEditorial score65
Remembering What Matters for Long-Horizon Robotic World Models
Remembering What Matters for Long-Horizon Robotic World Models
Summary
This research investigates the limitations of autoregressive robot video world models over long rollout horizons, proposing ReCAP (Receding Context with Anchor Preservation) as a training-free context-scheduling method. ReCAP preserves recent visual-action states and initial scene anchors while evict, improving prediction quality and computational efficiency across environments like BridgeV2 and CALVIN.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org