FRMD: Fast Robot Motion Diffusion with Consistency-Distilled Movement Primitives for Smooth Action Generation

Xirui Shi, Jun Jin

Keywords: Diffusion Models; Embodied AI

TL;DR: FRMD combines ProDMPs with consistency distillation to overcome diffusion models’ latency and smoothness issues, enabling real-time, one-step robot motion generation with state-of-the-art performance.

Abstract:

Diffusion models have recently emerged as a promising paradigm for robot action generation due to their scalability and ability to capture multimodal distributions [1], [5]. Despite this potential, their direct application to robotics is hindered by two key challenges. First, action-chunking variants generate short-horizon actions efficiently but fail to preserve temporal coherence, often resulting in jerky and unstable trajectories. Second, trajectory-level approaches such as Movement Primitive Diffusion (MPD) [2] leverage Probabilistic Dynamic Movement Primitives (ProDMPs) [3] to ensure smoothness, but require 10–50 iterative denoising steps, leading to prohibitive inference latency for real-time control.
To address this fundamental trade-off between efficiency and expressiveness, we introduce Fast Robot Motion Dif- fusion (FRMD), a framework that integrates ProDMPs with Consistency Models to enable smooth, structured, and tem- porally coherent motion generation in a single inference step. FRMD adopts a teacher–student distillation strategy, where a teacher diffusion model generates structured trajectories that are distilled into a student model capable of one-step infer- ence. Experimental evaluations on twelve manipulation tasks from the MetaWorld and ManiSkill benchmarks demonstrate that FRMD achieves state-of-the-art success rates while reducing inference time by an order of magnitude, thereby enabling real-time robotic control.

Submission Number: 69