Full-Process Adapted Diffusion Policy: Process-Level Joint Optimization for Robot Imitation Learning

Diffusion-based policies have shown strong performance in robot imitation learning, but directly adopting diffusion processes from image generation introduces mismatches with the low-dimensional, heterogeneous, and latency-sensitive nature of robot action spaces. To address this, we revisit the entire generative pipeline and propose the Full-process Adapted Diffusion Policy (FADP). Specifically, in the forward process, we analyze the widely used cosine noise schedule through a mode-separability framework, revealing its effectiveness in preserving mode discriminability in low-dimensional regimes and its interaction with subsequent processes. In the reverse process, we replace the shared inverse variance in standard DDPM with a learnable per-dimension formulation to better capture heterogeneous joint dynamics. In the sampling process, we employ consistency distillation to compress multi-step denoising into few-step inference, reducing latency while maintaining stability. These improvements require no architectural modifications and incur negligible overhead. Extensive experiments on both simulated and real-world robotic tasks demonstrate that FADP outperforms existing baselines and achieves competitive performance, highlighting the importance of full-process adaptation for diffusion-based robot policies.