Title: Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry

Abstract:

While diffusion-based and flow-matching generative action policies demonstrate superior performance in behavior cloning, their reliance on iterative sampling creates significant latency issues for high-frequency robotic control. Although recent one-step approaches mitigate this delay, they often sacrifice the intermediate trajectory evolution necessary for precise action correction. Reinstating this mechanism through the explicit estimation of a training-time drifting field is mathematically ill-posed, primarily due to the extreme sparsity of conditional demonstrations. To address this, we propose the Implicit Drifting Policy (IDP), a one-step imitation learning framework that integrates the corrective benefits of Drifting into policy learning without requiring explicit vector field estimation. IDP derives a conditional expert geometry by analyzing local variations in expert actions that are similar to current observations, then contrasts this with a global reference geometry to isolate constraints specific to the given condition. This local geometric structure dynamically weights a scalar potential objective. When paired with an expert-proximal terminal evaluation, IDP directly imposes manifold constraints on the one-step generator during the training phase. Comprehensive evaluations across 2D, 3D, and real-world manipulation tasks demonstrate that IDP effectively preserves adherence to valid action manifolds, outperforming explicit drifting methods and delivering competitive results against robust one-step baselines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC