Reproduce “AdaFlow: Imitation Learning with Variance-Adaptive Flow-Based Policies” 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
Xixi Hu, Bo Liu, Xingchao Liu, Qiang Liu
AdaFlow presents an imitation learning framework utilizing variance-adaptive flow-based policies to efficiently generate diverse, multi-modal actions. It achieves high task success rates in robotic manipulation and navigation, often requiring only 1-2 function evaluations during inference, significantly improving speed compared to prior generative approaches.
Key insights
- There is an intrinsic connection between the conditional variance of the training loss and the straightness of the learned ODE trajectories, where lower variance indicates straighter, simpler flows.
- Perfectly deterministic action distributions correspond to optimal ODE trajectories that are straight lines, allowing for accurate action generation in a single Euler step.
- The conditional variance can serve as a principled indicator to adaptively control the computational steps required for ODE solving, optimizing both inference speed and accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.