Title: Training-Free Imitation Learning with Closed-Form Diffusion Policies

Original: arXiv:2606.01238v1 Announce Type: cross Abstract: While diffusion-based policies have impressive performance and expressivity, their long offline training slows down the data collection and policy deployment loop. We introduce Closed-Form Diffusion Policies, a class of training-free diffusion-based policies for imitation learning using the closed-form score derived from the demonstration dataset. We deploy CFDP with real-time inference with a mobile CPU in hardware experiments, showing it can successfully perform imitation directly from the dataset in milliseconds and with faster inference than neural diffusion policies. In experiments on imitation learning benchmarks, we show that CFDP is competitive against neural baselines that require hours of training, providing a favorable tradeoff between training time and performance. Finally, we show how closed-form diffusion policies act as a composable primitive that enables data-driven inference-time editing of pre-trained neural diffusion policies, including policy guidance and novel demonstration augmentation.

Rewritten:

Title: Training-Free Imitation Learning with Closed-Form Diffusion Policies

Abstract: Although diffusion-based policies offer significant expressivity and high performance, the extensive offline training required often impedes the speed of the data collection and policy deployment cycle. To address this, we propose Closed-Form Diffusion Policies (CFDP), a novel category of training-free diffusion policies designed for imitation learning. These policies utilize a closed-form score calculated directly from the demonstration dataset. In hardware experiments involving real-time inference on a mobile CPU, we demonstrate that CFDP can execute imitation learning in milliseconds, achieving inference speeds that surpass those of neural diffusion policies. Furthermore, tests on standard imitation learning benchmarks reveal that CFDP remains competitive with neural baselines that necessitate hours of training, thereby offering an advantageous balance between computational efficiency and performance. Lastly, we illustrate how closed-form diffusion policies serve as a composable building block, facilitating data-driven, inference-time modifications to pre-trained neural diffusion policies, such as policy guidance and the augmentation of new demonstrations.

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