Diffusion Policy

The LeRobot ecosystem has built-in support for imitation learning algorithms such as ACT and Diffusion Policy, making it possible to run the pipeline from teleoperation data collection to policy training within a few days.

现代视觉-语言-动作（VLA）策略（如 ACT、Diffusion Policy、π0）不再逐步查询动作，而是一次前向传播生成未来 H 步的动作序列（chunk），再开环执行其中长度为 L 的前缀后重新查询。

The key innovation of ACT lies in action chunking, which packs multiple consecutive action steps into a single action sequence chunk for unified prediction, rather than outputting single-step actions frame by frame.

Implicit behavior cloning and diffusion policies alleviate the averaging behavior caused by distribution shift by modeling the multimodal distribution of actions rather than regressing to the mean.

Toyota Research Institute's Diffusion Policy and Physical Intelligence's π0 demonstrate that using a diffusion model as an action generator is more suitable for continuous control tasks than autoregressive token prediction.

Google's RT series models, the Toyota Research Institute's Diffusion Policy, and Vision-Language-Action (VLA) models are early attempts toward world task models.

Motion planning algorithms are transitioning from handcrafted methods (RRT, A*, CHOMP, TrajOpt) toward data-driven approaches (such as Diffusion Policy).

The core advantage of diffusion policy lies in its ability to model multimodal action distributions—when a single observed state corresponds to multiple reasonable actions, it can sample coherent action sequences, whereas traditional regression methods output averaged actions and thus fail.

Diffusion policy has demonstrated performance superior to traditional behavioral cloning and Transformer-based policies on multiple real robot tasks.

Diffusion Policy is a method proposed in 2023 by a team from MIT and Columbia University, drawing on Denoising Diffusion Probabilistic Models (DDPM).