How do we achieve these surprising results? Instead of learning a global, direct state-to-action mapping, DARP reparameterizes the imitation learning problem through retrieval.

We retrieve the k-nearest expert neighbors and predict actions conditioned on the relative difference vectors between those neighbor states and the query state, aggregating them for a final action prediction. (2/7)

But why does DARP provide such large gains just through reparameterization? Theoretically, we show that DARP operates as an implicit manifold regularizer. By embedding neighborhood aggregation directly into the policy architecture, it achieves parameter-free Laplacian smoothing. This causes smooth, low-variance behavior, leading to more stable models and improved performance. (3/7)

Crucially, DARP remains entirely within the standard behavior cloning regime: no simulators, interactive experts, or online training are required.

DARP sees substantial improvement across architectures and input modalities, and even significantly improves upon diffusion policies in real-world robotic manipulation tasks. (5/7)

Empirically, DARP is dead simple to implement, just a little nearest neighbors + the simple aggregation architecture - no objective change, no data change. Since DARP is easy to scale, you can drop it in for your favorite policy class! Nicely integrates with rich distribution classes like diffusion and works directly from visual inputs! You basically get pretty huge gains across the board with very few changes required. (6/7)

Empirically, this difference-based reparameterization directly mitigates covariate shift.

By shifting the frame of reference to local distance vectors, query states that are globally out-of-distribution (OOD) remain locally in-distribution. This local consistency grants the policy remarkable robustness during closed-loop rollouts. (4/7)