FODMP: Fast One-Step Diffusion of Movement Primitives Generation for Time-Dependent Robot Actions
New AI framework distills diffusion models for 10x faster inference, letting robots catch fast-flying balls in real time.
A research team has introduced FODMP (Fast One-step Diffusion of Movement Primitives), a novel framework that overcomes a critical bottleneck in AI-powered robot control. Current methods face a harsh trade-off: action-chunking diffusion policies (like ManiCM) are fast but only predict short, reactive motion segments, while Movement Primitive Diffusion (MPD) can generate full, temporally-structured trajectories but suffers from prohibitively high inference latency due to its multi-step process. FODMP solves this by distilling the diffusion model into the parameter space of Probabilistic Dynamic Movement Primitives (ProDMPs), allowing it to generate complex, time-dependent motions—complete with acceleration and deceleration profiles—using a single-step decoder.
This architectural breakthrough translates to dramatic performance gains. On standard benchmarks like MetaWorld and ManiSkill, FODMP runs up to 10 times faster than MPD and 7 times faster than action-chunking policies, while matching or exceeding their task success rates. More importantly, the speed enables new capabilities in dynamic, real-world environments. The paper demonstrates that FODMP allows a robot to visually track and securely catch a fast-flying ball in real time, a task where both previous categories of models responded too slowly to even attempt an interception. This marks a significant step toward deploying learned, temporally-aware robot policies in settings requiring high-frequency, closed-loop control.
- Uses consistency distillation to generate ProDMP-based motion primitives in a single decoding step, eliminating multi-step diffusion latency.
- Achieves up to 10x faster inference than MPD and 7x faster than action-chunking policies on MetaWorld and ManiSkill benchmarks.
- Enables real-time, vision-based dynamic tasks like intercepting a fast-flying ball, which was infeasible with prior slower models.
Why It Matters
Bridges the gap between high-quality motion generation and real-time control, enabling robots to perform dynamic, time-sensitive tasks in unstructured environments.