Reproduce “BWM: A Low-Cost High-Fidelity World Simulator for Robot Learning” 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
BWM Team
Researchers introduce BWM, an open-source, action-conditioned world simulator for robot manipulation that achieves state-of-the-art fidelity and low training costs by adapting a pretrained video diffusion model. The simulator ranked first in the WorldArena Challenge with an EWMScore of 63.51 and improved physical robot policy success rates by over 20% when used for data augmentation.
Key insights
- A 'domain-specific robot post-training' strategy on pretrained video diffusion models provides an efficient and effective route to high-fidelity robot world simulation at a lower computational cost.
- Precise action conditioning is best achieved by injecting robot actions through both frame-level cross-attention for direct cues and latent-level AdaLN for broader modulation of the denoising process.
- Combining initial-environment guidance with a dynamic history window is crucial for maintaining overall scene consistency and accurately tracking visual changes over long, autoregressive rollouts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.