ManiCM: Consistency Model Enables Real-Time 3D Robotic Manipulation

ManiCM, a novel robotic manipulation model, uses a consistency model to generate robot actions in a single inference step, overcoming the runtime inefficiency of diffusion-based approaches. The model applies a consistency constraint to the diffusion process, allowing direct denoising from any point along the ODE trajectory within the robot action space, conditioned on point cloud observations. A consistency distillation technique predicts the action sample directly instead of the noise. Diffusion models are effective at generating complex distributions, and recent diffusion-based methods excel at 3D robotic manipulation but require multiple denoising steps, especially with high-dimensional observations. ManiCM addresses this by imposing the consistency constraint, enabling real-time performance.