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

Researchers introduce Implicit Drifting Policy (IDP), a one-step imitation learning framework that enables faster robot control by extracting conditional expert geometry from demonstration data rather than explicitly estimating drift fields. IDP maintains adherence to valid action manifolds while achieving competitive performance with existing methods across manipulation tasks.

IDP addresses a fundamental challenge in robotic control: the tension between sample quality and computational speed. Diffusion-based and flow-matching policies excel at behavior cloning through iterative refinement, but their multi-step sampling creates latency unsuitable for real-time robot control applications. Recent one-step approaches accelerate inference dramatically but sacrifice the intermediate trajectory corrections that ensure action validity, leading to decreased performance.

The core innovation lies in how IDP recovers correction mechanisms without explicit drift field estimation, which proves mathematically intractable given sparse conditional demonstrations. Instead, the method extracts geometric structure from local variations in expert actions conditioned on similar observations, then compares this against global reference geometry to isolate constraint-specific patterns. This implicit geometric reasoning adaptively weights the training objective, enforcing manifold constraints during learning rather than at inference.

The approach demonstrates broad applicability across 2D simulations, 3D environments, and physical manipulation tasks, consistently maintaining action validity while matching or exceeding baseline performance. For robotics practitioners, this represents a meaningful advance in deploying learned policies in high-frequency control scenarios where latency directly impacts task success. The geometric interpretation also provides interpretability advantages, allowing practitioners to understand which constraints the policy learns from demonstration data.

Future development should explore whether this implicit geometry extraction scales to higher-dimensional action spaces and whether the method generalizes to domains with greater demonstration diversity or longer horizon tasks requiring sequential decision-making.

• →IDP enables one-step robot control policy generation by implicitly learning correction mechanisms through conditional expert geometry extraction.
• →The method avoids ill-posed explicit drift field estimation by comparing local observation-conditioned action variation against global reference geometry.
• →IDP maintains adherence to valid action manifolds while achieving competitive performance with existing one-step and iterative baselines.
• →The geometric framework provides adaptive weighting of training objectives, directly enforcing constraints during learning rather than inference.
• →Experimental validation spans 2D, 3D, and real-world manipulation tasks, demonstrating broad applicability for practical robotic control.