Diffusion policy: Visuomotor policy learning via action diffusion

By Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, Shuran Song. Robotics: Science and Systems Conference, 2023

3,739 citations recorded by the source corpus.

This paper introduces Diffusion Policy, a new way of generating robot behavior by representing a robot’s visuomotor policy as a conditional denoising diffusion process. We benchmark Diffusion Policy across 15 different tasks from 4 different robot manipulation benchmarks and find that it consistently outperforms existing state-of-the-art robot learning methods with an average improvement of 46.9%. Diffusion Policy learns the gradient of the action-distribution score function and iteratively optimizes with respect to this gradient field during inference via a series of stochastic Langevin dynamics steps. We find that the diffusion formulation yields powerful advantages when used for robot policies, including gracefully handling multimodal action distributions, being suitable for high-dimensional action spaces, and exhibiting impressive training stability. To fully unlock the potential of diffusion models for visuomotor policy learning on physical robots, this paper presents a set of key technical contributions including the incorporation of receding horizon control, visual conditioning, and the time-series diffusion transformer. We hope this work will help motivate a new generation of policy learning techniques that are able to leverage the powerful generative modeling capabilities of diffusion models. Code, data, and training details are available (diffusion-policy.cs.columbia.edu).

Subfields: Artificial Intelligence

View Paper: https://www.semanticscholar.org/paper/bdba3bd30a49ea4c5b20b43dbd8f0eb59e9d80e2

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

π0: A Vision-Language-Action Flow Model for General Robot Control (2024, neutral): “We also compare to ACT [57] and Diffusion Policy [9], which are designed specifically for learning dexterous tasks from smaller datasets.”
CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (2025, positive): “Diffusion Policy [8], a state-of-the-art imitation learning algorithm, is trained from scratch for each test scenario in LIBERO and Franka-Tabletop.”
RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots (2024, background): “Wealso experiment with diffusion policy [4].”
TinyVLA: Toward Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation (2024, neutral): “Moreover, traditional robot models struggle to adapt to new scenes and tasks and are easily affected by distractors, lighting conditions, and background changes [3], [4].”
RVT: Robotic View Transformer for 3D Object Manipulation (2023, background): “The success of transformers in vision and NLP has led its way into robot learning [42, 43, 44, 17, 45, 46].”
SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model (2025, neutral): “XIV present comprehensive evaluation results comparing the performance of four state-of-the-art approaches: Diffusion Policy [12], Octo [48], Open-VLA [30], and our proposed SpatialVLA on the Franka Panda Emika robot.”
Autoregressive Image Generation without Vector Quantization (2024, background): “Our work is conceptually related to Diffusion Policy [8] in robotics.”
Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation (2023, methodology): “2022; Ahn et al., 2022; Driess et al., 2023; Du et al., 2023; Guhur et al., 2023; Jang et al., 2022; Lynch & Sermanet, 2020; Lynch et al., 2022; Mees et al., 2022a;c; Ha et al., 2023).”
Open-TeleVision: Teleoperation with Immersive Active Visual Feedback (2024, neutral): “If we distinguish the existing work with sources of demonstrations, we can classify them into learning from real-robot expert data [14, 10, 12, 59, 60, 61, 62, 50, 63, 1, 64, 51, 65], learning from play data [66, 67, 68], and learning from human demonstrations [67, 69, 70, 71, 72, 73, 74, 56].”
NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration (2023, methodology): “The noise prediction network, ϵ θ , consists of a 1D conditional U-Net [29, 31] with 15 convolutional layers.”
OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning (2024, methodology): “We benchmark a variety of imitation learning algorithms on four tasks in our collected dataset (shown in Figure 7), including Diffusion Policy [58] with Denoising Diffusion Probabilistic Model [59] (DP-DDPM) and Denoising Diffusion Implicit Model [60] (DP-DDIM) and vanilla Behavior Cloning with a…”
Evaluating Real-World Robot Manipulation Policies in Simulation (2024, neutral): “Much recent work in robotic manipulation has demonstrated impressive tasks that go beyond rigid object manipulation [20, 10, 33].”
One Step Diffusion via Shortcut Models (2024, neutral): “We compare to iterative denoising meth-ods IBC (Florence et al., 2022) and diffusion policy (Chi et al., 2023), as well as one-step methods LSTM-GMM (Mandlekar et al., 2021) and BET (Shafiullah et al., 2022).”
A Survey on Vision--Language--Action Models for Embodied AI (2024, neutral): “…[43] and VIMA [150] use SE (2) action and its BC objective can be expressed as follows: The DDPM objective in diffusion-based control policies [130] is represented as: VLA However, more complex tasks—such as “pouring water into a cup”—may require additional DoFs, thereby necessitating SE (3)…”
DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models (2023, background): “These models have proven to be highly effective in a range of domains, including image generation [4], audio generation [14], 3D synthesis [26], and robotics [3].”
Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Models (2023, methodology): “We use a diffusion model for our goal-reaching policy since recent work has shown that diffusion-based policies can better capture multi-modality in robot data [12, 23], leading to improved performance across a variety of tasks.”
DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset (2024, positive): “DROID’s interactions cover a larger part of the robot’s workspace, since the robot is moved freely between collection sessions instead of being placed in front of repetitive tabletop scenes. a wide spectrum of robot manipulation tasks and environments.”
RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (2024, positive): “Besides, according to Chi et al. (2023), action chunking is also helpful for improving temporal consistency.”
Octo: An Open-Source Generalist Robot Policy (2024, methodology): “It also supports natural language instructions, goal images, observation histories, and multi-modal, chunked action prediction via diffusion decoding [17].”
CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation (2024, neutral): “However, feasible actions for task execution can belong to different modes [15], and simply ag-gregating them could result in an action that does not align with any mode, which is suboptimal.”