Skip to main content

Diffusion Policy — Visuomotor Control via Denoising

Duration: 60 min · Level: Advanced · Module: 5. Foundation Models & VLA Architecture · Focus: diffusion-policy, visuomotor, bimanual, manipulation

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • Key insight
  • Two variants
  • Architecture
  • State-of-the-art on
  • ACT (Action Chunking with Transformers, Zhao et al. 2023)

Why this matters​

Diffusion Policy (Chi et al., RSS 2023) applies denoising diffusion probabilistic models to robot action generation.

Overview​

Diffusion Policy (Chi et al., RSS 2023) applies denoising diffusion probabilistic models to robot action generation. Unlike direct regression, it learns a score function that can represent multimodal action distributions — critical for tasks where multiple correct trajectories exist (e.g., "place the cup somewhere on the table").

Key concepts​

Key idea

Key insight: direct regression predicts a single mean action; diffusion policy predicts a distribution; crucial for contact-rich tasks with multiple valid solutions

  • Two variants: DDPM-based (slower, 100-step denoising) and consistency policy (faster, 1-4 step denoising); for real-time use, consistency policy preferred
  • Architecture: conditioning on RGB observations via ResNet/ViT encoder; denoising network is U-Net or transformer; action space is joint angles or end-effector deltas
  • State-of-the-art on: RoboMimic, RLBench, BridgeV2; particularly strong on bimanual tasks and contact-rich manipulation
  • ACT (Action Chunking with Transformers, Zhao et al. 2023): related approach using chunked action prediction; used in Stanford ALOHA bimanual robot system
  • ALOHA 2 (Stanford, 2024): 14-DOF bimanual system trained with ACT on 50 demonstrations; achieves cooking, surgery simulation tasks with 90%+ success rate

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about Key insight?

direct regression predicts a single mean action; diffusion policy predicts a distribution; crucial for contact-rich tasks with multiple valid solutions

Q2. What do you know about Two variants?

DDPM-based (slower, 100-step denoising) and consistency policy (faster, 1-4 step denoising); for real-time use, consistency policy preferred

Q3. What do you know about Architecture?

conditioning on RGB observations via ResNet/ViT encoder; denoising network is U-Net or transformer; action space is joint angles or end-effector deltas

Q4. What do you know about State-of-the-art on?

RoboMimic, RLBench, BridgeV2; particularly strong on bimanual tasks and contact-rich manipulation

Q5. What do you know about ACT (Action Chunking with Transformers, Zhao et al. 2023)?

related approach using chunked action prediction; used in Stanford ALOHA bimanual robot system

References​

  • Diffusion Policy: Visuomotor Policy Learning via Action Diffusion — Chi et al. (2023). RSS 2023
  • Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware — Zhao et al. (2023). RSS 2023

← Previous: 5.3 OpenVLA — The Open-Source VLA Ecosystem · Next: 5.5 Deploying VLAs on G1: Architecture & Integration →

Part of Module 5: Foundation Models & VLA Architecture.