Reproduce “L1 Sample Flow for Efficient Visuomotor 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
Weixi Song, Zhetao Chen, Tao Xu, Xianchao Zeng, Xinyu Zhou, Lixin Yang, Donglin Wang, Cewu Lu and 1 more
L1 Flow reconfigures flow matching for visuomotor policy learning, enabling robust multi-modal action distribution modeling with significantly reduced inference latency and faster training convergence. This approach achieves comparable or superior task performance while demonstrating 10-70x faster inference than traditional diffusion policies and 5-7x faster training convergence than distillation methods.
Key insights
- Flow matching can be effectively re-framed to directly predict action samples rather than velocities, allowing for the integration of efficient L1 regression objectives.
- A single ODE integration step is empirically sufficient to capture the multi-modal nature of target action distributions, negating the need for numerous iterative denoising steps.
- L1 loss consistently outperforms MSE loss for the sample-prediction flow matching variant, leading to improved performance in visuomotor control tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.