Reproduce “Rethink Before You Execute: Adaptive Execution for World Action Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Feng Ye, Yiming Zhao, Yong Yu, Hongxu Zhou, Yong Pan, Yuan Xue, Peng Jia, Chuanmin Jia
TempoWAM adds a progress-based execution rule to World Action Models (WAMs), which predict robot action bundles. It estimates task progress from observations and execution history, continuing a bundle when progress is sufficient and otherwise replanning. On real robots, it keeps “take drinks” success at 90.0% with 26.9% fewer calls, but “pack hand cream” success rises to 63.3% as calls increase.
Key insights
- Dynamically adjusting the execution horizon based on explicit task progress directly optimizes the trade-off between execution efficiency and task success in robot manipulation.
- A lightweight, recurrent monitor can effectively estimate task progress with minimal computational overhead, allowing for adaptive replanning decisions without requiring extensive WAM re-inference.
- A sophisticated two-timescale online calibration protocol (intra-episode EMA and inter-episode success-gated adaptation) enables robust, self-tuning adaptation to varying task conditions and performance targets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.