Reproduce “WoW: Towards a World omniscient World model Through Embodied Interaction” 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
Xiaowei Chi, Peidong Jia, Chun-Kai Fan, Xiaozhu Ju, Weishi Mi, Kevin Zhang, Zhiyuan Qin, Wanxin Tian and 28 more
The WoW project introduces a generative world model, grounded in 7,300 hours of robot interaction data, that develops authentic physical intuition by integrating perception, imagination, reflection, and action. It achieves state-of-the-art performance on a new benchmark, WoWBench, and demonstrates robust real-world robot manipulation.
Key insights
- Authentic physical intuition for world models is fundamentally grounded in extensive, causally rich interactions with the real world, rather than solely passive observational data.
- The SOPHIA 'Prover-Verifier' paradigm effectively instills abstract values like physical common sense into generative models through iterative self-refinement and VLM-guided feedback.
- Systematic scaling analyses reveal that performance of embodied world models improves significantly with increased data volume and model size, indicating continued potential for growth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.