FRMD: fast robot motion diffusion via trajectory-level consistency distillation.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Foundation models for Embodied AI often use diffusion modules for action generation.
• The iterative denoising process in diffusion modules leads to high inference latency, hindering real-time deployment.
• Existing Vision-Language-Action (VLA) policies face challenges with real-time performance due to diffusion module bottlenecks.

Purpose of the Study:

• To address the inference latency bottleneck in diffusion-based action generation for Embodied AI.
• To develop a fast robot motion diffusion framework (FRMD) for real-time robotic control.
• To enable practical, real-time applications of VLA policies by optimizing the motion generation component.

Main Methods:

• FRMD operates in trajectory-parameter space, predicting movement-primitive coefficients on a low-dimensional manifold.
• A novel trajectory-level consistency distillation technique collapses multi-step sampling into a single inference step.
• The framework replaces stepwise action generation with a one-pass mapping from noise to full trajectories, followed by basis expansion.

Main Results:

• FRMD reduces policy latency from hundreds to tens of milliseconds without altering upstream vision or language encoders.
• Achieved 7x speedup compared to vanilla diffusion policies and 10x speedup over the state-of-the-art MPD method.
• Maintained task success rates comparable to multi-step diffusion policies on robotic manipulation benchmarks.

Conclusions:

• FRMD offers a plug-in, latency-optimized motion generator for VLA systems.
• The framework preserves the benefits of diffusion models while making real-time Embodied AI feasible.
• This advancement is crucial for deploying sophisticated AI in real-world robotic applications requiring immediate responses.