Diffusion-based action policy implementation.

Implements the generative action policy that produces candidate action sequences given observational input, using DDPM (Denoising Diffusion Probabilistic Models).

Architecture

The policy uses a conditional denoising network that takes:

• Noisy action sequence [batch, pred_horizon * action_dim]
• Observation conditioning [batch, obs_horizon * obs_dim]
• Diffusion timestep embedding

And predicts the noise to remove, iteratively denoising pure Gaussian noise into a valid action sequence.