Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 86 of the 86 references it cites were found in a registry or an index; 16 of the 39 sources its claims cite were checked for support.

Diffusion policy: Visuomotor policy learning via action diffusion

Chi, Cheng, Xu, Zhenjia, Feng, Siyuan, Cousineau, Eric, Du, Yilun, Burchfiel, Benjamin, Tedrake, Russ, Song, Shuran · 2025 · The International Journal of Robotics Research

Abstract

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).

Overview

CoChat Summary

TL;DR

Diffusion Policy represents a visuomotor policy as a conditional denoising diffusion process over action sequences, using receding-horizon control, visual conditioning, and CNN or transformer backbones. Across 15 tasks from 4 benchmarks it reports an average 46.9% improvement over prior state-of-the-art, with strong real-world manipulation results.

Key findings

  • Across 15 tasks from 4 robot manipulation benchmarks, Diffusion Policy consistently outperforms prior state-of-the-art with an average improvement of 46.9%; in simulation it reports best/average success rates such as 1.00/1.00 on Lift ph and 1.00/0.87 on ToolHang ph for DiffusionPolicy-T (Table 1).
  • Real-world Push-T: 95% success and 0.8 average IoU vs human 1.00 success and 0.84 IoU; best IBC and LSTM-GMM variants reached 0% and 20% success.
  • Other real-world tasks: Mug Flipping 90% success over 20 trials (LSTM-GMM 0%); Sauce Pouring 0.74 IoU / 0.79 success vs human 0.79 / 1.00; Sauce Spreading 0.77 coverage / 1.00 success vs human 0.79 / 1.00; bimanual Egg Beater 55%, Mat Unrolling 75%, Shirt Folding 75%, each over 20 trials.
  • Ablations: action horizon of 8 steps was optimal for most tested tasks; Diffusion Policy maintained peak performance with latency up to 4 steps; position control outperformed velocity control; DDIM with 100 training and 10 inference iterations enabled 0.1 s inference latency on an Nvidia 3080 GPU.

Methods

  • Design: conditional DDPM/DDIM policy \(p(A_t|O_t)\) that predicts action sequences; uses receding-horizon control, visual conditioning, and either a CNN with FiLM or a time-series diffusion transformer.
  • Data/evaluation: 15 tasks from 4 benchmarks (Robomimic, Push-T, Block Push, Kitchen) plus real-world single-arm and bimanual tasks; simulation results averaged over 3 training seeds and 50 environment initial conditions (1500 experiments total; robomimic used 22 initializations due to an evaluation-code bug).
  • Analysis: compares against LSTM-GMM, IBC, and BET; metrics include success rate, target-area coverage/IoU, and ablations on action horizon, latency, action space, vision encoder, and architecture.

Limitations

  • Stated: inherits behavior cloning limitations such as suboptimal performance with inadequate demonstration data; higher computational cost and inference latency than simpler methods like LSTM-GMM, which action-sequence prediction only partially mitigates and may not suffice for high-rate control.
  • Stated: transformer backbone is more sensitive to hyperparameters; CNN backbone performs poorly when action sequences change quickly and sharply through time, such as velocity command action space.
  • Evident: real-world evaluations often use small trial counts (e.g., 20 trials per task), and bimanual success rates were 55–75%; simulation robomimic used only 22 initializations due to a bug; frozen pretrained vision encoders performed poorly, and ViT from scratch reached only 22% success on square.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1177/02783649241273668

Cites

86 of 86 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

p(a) Figure 1. Policy Representations. a) Explicit policy with different types of action representations. b) Implicit policy learns an energy function conditioned on both action and observation and optimizes for actions that minimize the energy landscape c) Diffusion policy refines noise into actions via a learned gradient field. This formulation provides stable training, allows the learned policy to accurately model multimodal action distributions, and accommodates high-dimensional action sequences.
p(a) Figure 1. Policy Representations. a) Explicit policy with different types of action representations. b) Implicit policy learns an energy function conditioned on both action and observation and optimizes for actions that minimize the energy landscape c) Diffusion policy refines noise into actions via a learned gradient field. This formulation provides stable training, allows the learned policy to accurately model multimodal action distributions, and accommodates high-dimensional action sequences.From the PDF
Figure 2 Diffusion Policy Overview a) General formulation. At time step t , the policy takes the latest To steps of observation data Ot as input and outputs Ta steps of actions At . b) In the CNN-based Diffusion Policy, FiLM (Feature-wise Linear Modulation) Perez et al. (2018) conditioning of the observation feature Ot is applied to every convolution layer, channel-wise. Starting from A K t drawn from Gaussian noise, the output of noise-prediction network εθ is subtracted, repeating K times to get A 0 t , the denoised action sequence. c) In the Transformer-based Vaswani et al. (2017) Diffusion Policy, the embedding of observation O t is passed into a multi-head cross-attention layer of each transformer decoder block. Each action embedding is constrained to only attend to itself and previous action embeddings (causal attention) using the attention mask illustrated.
Figure 2 Diffusion Policy Overview a) General formulation. At time step t , the policy takes the latest To steps of observation data Ot as input and outputs Ta steps of actions At . b) In the CNN-based Diffusion Policy, FiLM (Feature-wise Linear Modulation) Perez et al. (2018) conditioning of the observation feature Ot is applied to every convolution layer, channel-wise. Starting from A K t drawn from Gaussian noise, the output of noise-prediction network εθ is subtracted, repeating K times to get A 0 t , the denoised action sequence. c) In the Transformer-based Vaswani et al. (2017) Diffusion Policy, the embedding of observation O t is passed into a multi-head cross-attention layer of each transformer decoder block. Each action embedding is constrained to only attend to itself and previous action embeddings (causal attention) using the attention mask illustrated.From the PDF
Figure 3 Multimodal behavior. At the given state, the end-effector (blue) can either go left or right to push the block. Diffusion Policy learns both modes and commits to only one mode within each rollout. In contrast, both LSTM-GMM Mandlekar et al. (2021) and IBC Florence et al. (2021) are biased toward one mode, while BET Shafiullah et al. (2022) fails to commit to a single mode due to its lack of temporal action consistency. Actions generated by rolling out 40 steps for the best-performing checkpoint.
Figure 3 Multimodal behavior. At the given state, the end-effector (blue) can either go left or right to push the block. Diffusion Policy learns both modes and commits to only one mode within each rollout. In contrast, both LSTM-GMM Mandlekar et al. (2021) and IBC Florence et al. (2021) are biased toward one mode, while BET Shafiullah et al. (2022) fails to commit to a single mode due to its lack of temporal action consistency. Actions generated by rolling out 40 steps for the best-performing checkpoint.From the PDF
Figure 4 Velocity v.s. Position Control. The performance difference when switching from velocity to position control. While both BCRNN and BET performance decrease, Diffusion Policy is able to leverage the advantage of position and improve its performance.
Figure 4 Velocity v.s. Position Control. The performance difference when switching from velocity to position control. While both BCRNN and BET performance decrease, Diffusion Policy is able to leverage the advantage of position and improve its performance.From the PDF
Figure 5 Diffusion Policy Ablation Study. Change (difference) in success rate relative to the maximum for each task is shown on the Y-axis. Left : trade-off between temporal consistency and responsiveness when selecting the action horizon. Right : Diffusion Policy with position control is robust against latency. Latency is defined as the number of steps between the last frame of observations to the first action that can be executed.
Figure 5 Diffusion Policy Ablation Study. Change (difference) in success rate relative to the maximum for each task is shown on the Y-axis. Left : trade-off between temporal consistency and responsiveness when selecting the action horizon. Right : Diffusion Policy with position control is robust against latency. Latency is defined as the number of steps between the last frame of observations to the first action that can be executed.From the PDF
Figure 6 Training Stability. Left: IBC fails to infer training actions with increasing accuracy despite smoothly decreasing training loss for energy function. Right: IBC's evaluation success rate oscillates, making checkpoint selection difficult (evaluated using policy rollouts in simulation).
Figure 6 Training Stability. Left: IBC fails to infer training actions with increasing accuracy despite smoothly decreasing training loss for energy function. Right: IBC's evaluation success rate oscillates, making checkpoint selection difficult (evaluated using policy rollouts in simulation).From the PDF
Image 7
From the PDF
Image 8
From the PDF
Image 9
From the PDF
Image 10
From the PDF
Image 11
From the PDF
Table 6. Realworld Push-T Experiment. a) Hardware setup. b) Illustration of the task. The robot needs to 1 ⃝ precisely push the T-shaped block into the target region, and 2 ⃝ move the end-effector to the end-zone. c) The ground truth end state used to calculate IoU metrics used in this table. Table: Success is defined by the end-state IoU greater than the minimum IoU in the demonstration dataset. Average episode duration presented in seconds. T-E2E stands for end-to-end trained Transformer-based Diffusion Policy
Table 6. Realworld Push-T Experiment. a) Hardware setup. b) Illustration of the task. The robot needs to 1 ⃝ precisely push the T-shaped block into the target region, and 2 ⃝ move the end-effector to the end-zone. c) The ground truth end state used to calculate IoU metrics used in this table. Table: Success is defined by the end-state IoU greater than the minimum IoU in the demonstration dataset. Average episode duration presented in seconds. T-E2E stands for end-to-end trained Transformer-based Diffusion PolicyFrom the PDF
Figure 7 Realworld Push-T Comparisons. Columns 1-4 show action trajectories based on key events. The last column shows averaged images of the end state. A : Diffusion policy (End2End) achieves more accurate and consistent end states. B : Diffusion Policy (R3M) gets stuck initially but later recovers and finishes the task. C : LSTM-GMM fails to reach the end zone while adjusting the T block, blocking the eval camera view. D : IBC prematurely ends the pushing stage.
Figure 7 Realworld Push-T Comparisons. Columns 1-4 show action trajectories based on key events. The last column shows averaged images of the end state. A : Diffusion policy (End2End) achieves more accurate and consistent end states. B : Diffusion Policy (R3M) gets stuck initially but later recovers and finishes the task. C : LSTM-GMM fails to reach the end zone while adjusting the T block, blocking the eval camera view. D : IBC prematurely ends the pushing stage.From the PDF
Figure 8 Robustness Test for Diffusion Policy. Left : A waving hand in front of the camera for 3 seconds causes slight jitter, but the predicted actions still function as expected. Middle : Diffusion Policy immediately corrects shifted block position to the goal state during the pushing stage. Right : Policy immediately aborts heading to the end zone, returning the block to goal state upon detecting block shift. This novel behavior was never demonstrated. Please check the videos in the supplementary material.
Figure 8 Robustness Test for Diffusion Policy. Left : A waving hand in front of the camera for 3 seconds causes slight jitter, but the predicted actions still function as expected. Middle : Diffusion Policy immediately corrects shifted block position to the goal state during the pushing stage. Right : Policy immediately aborts heading to the end zone, returning the block to goal state upon detecting block shift. This novel behavior was never demonstrated. Please check the videos in the supplementary material.From the PDF
Figure 9 6DoF Mug Flipping Task. The robot needs to 1 ⃝ Pickup a randomly placed mug and place it lip down (marked orange). 2 ⃝ Rotate the mug such that its handle is pointing left.
Figure 9 6DoF Mug Flipping Task. The robot needs to 1 ⃝ Pickup a randomly placed mug and place it lip down (marked orange). 2 ⃝ Rotate the mug such that its handle is pointing left.From the PDF
Figure 10 Realworld Sauce Manipulation. [Left] 6DoF pouring Task. The robot needs to 1 ⃝ dip the ladle to scoop sauce from the bowl, 2 ⃝ approach the center of the pizza dough, 3 ⃝ pour sauce, and 4 ⃝ lift the ladle to finish the task. [Right] Periodic spreading Task The robot needs to 1 ⃝ approach the center of the sauce with a grasped spoon, 2 ⃝ spread the sauce to cover pizza in a spiral pattern, and 3 ⃝ lift the spoon to finish the task.
Figure 10 Realworld Sauce Manipulation. [Left] 6DoF pouring Task. The robot needs to 1 ⃝ dip the ladle to scoop sauce from the bowl, 2 ⃝ approach the center of the pizza dough, 3 ⃝ pour sauce, and 4 ⃝ lift the ladle to finish the task. [Right] Periodic spreading Task The robot needs to 1 ⃝ approach the center of the sauce with a grasped spoon, 2 ⃝ spread the sauce to cover pizza in a spiral pattern, and 3 ⃝ lift the spoon to finish the task.From the PDF
Figure 11 Bimanual Egg Beater Manipulation. The robot needs to 1 ⃝ push the bowl into position (only if too close to the left arm), 2 ⃝ approach and pick up the egg beater with the right arm, 3 ⃝ place the egg beater in the bowl, 4 ⃝ approach and grasp the egg beater crank handle, and 5 ⃝ turn the crank handle 3 or more times.
Figure 11 Bimanual Egg Beater Manipulation. The robot needs to 1 ⃝ push the bowl into position (only if too close to the left arm), 2 ⃝ approach and pick up the egg beater with the right arm, 3 ⃝ place the egg beater in the bowl, 4 ⃝ approach and grasp the egg beater crank handle, and 5 ⃝ turn the crank handle 3 or more times.From the PDF
Figure 12 Bimanual Mat Unrolling. The robot needs to 1 ⃝ pick up one side of the mat (if needed), using the left or right arm, 2 ⃝ lift and unroll the mat (if needed), 3 ⃝ ensure that both sides of the mat are grasped, 4 ⃝ lift the mat, 5 ⃝ place the mat oriented with the table, mostly centered, and 6 ⃝ release the mat.
Figure 12 Bimanual Mat Unrolling. The robot needs to 1 ⃝ pick up one side of the mat (if needed), using the left or right arm, 2 ⃝ lift and unroll the mat (if needed), 3 ⃝ ensure that both sides of the mat are grasped, 4 ⃝ lift the mat, 5 ⃝ place the mat oriented with the table, mostly centered, and 6 ⃝ release the mat.From the PDF
Figure 13 Bimanual Shirt Folding. The robot needs to 1 ⃝ approach and grasp the closest sleeve with both arms, 2 ⃝ fold the sleeve and release, 3 ⃝ drag the shirt closer (if needed), 4 ⃝ approach and grasp the other sleeve with both arms, 5 ⃝ fold the sleeve and release, 6 ⃝ drag the shirt to a orientation for folding, 7 ⃝ grasp and fold the shirt in half by its collar, 8 ⃝ drag the shirt to the center, and 9 ⃝ smooth out the shirt and move the arms away.
Figure 13 Bimanual Shirt Folding. The robot needs to 1 ⃝ approach and grasp the closest sleeve with both arms, 2 ⃝ fold the sleeve and release, 3 ⃝ drag the shirt closer (if needed), 4 ⃝ approach and grasp the other sleeve with both arms, 5 ⃝ fold the sleeve and release, 6 ⃝ drag the shirt to a orientation for folding, 7 ⃝ grasp and fold the shirt in half by its collar, 8 ⃝ drag the shirt to the center, and 9 ⃝ smooth out the shirt and move the arms away.From the PDF
Figure 14 Observation Horizon Ablation Study. State-based Diffusion Policy is not sensitive to observation horizon. Vision-based Diffusion Policy prefers low but > 1 observation horizon, with 2 being a good compromise for most tasks.
Figure 14 Observation Horizon Ablation Study. State-based Diffusion Policy is not sensitive to observation horizon. Vision-based Diffusion Policy prefers low but > 1 observation horizon, with 2 being a good compromise for most tasks.From the PDF
Figure 15 Data Efficiency Ablation Study. Diffusion Policy outperforms LSTM-GMM Mandlekar et al. (2021) at every training dataset size.
Figure 15 Data Efficiency Ablation Study. Diffusion Policy outperforms LSTM-GMM Mandlekar et al. (2021) at every training dataset size.From the PDF
Figure 16 Initial states for Mat Unrolling
Figure 16 Initial states for Mat UnrollingFrom the PDF
Figure 17 Final states for Mat Unrolling
Figure 17 Final states for Mat UnrollingFrom the PDF
Image 24
From the PDF
Image 25
From the PDF
Figure 20 Initial states for Shirt Folding
Figure 20 Initial states for Shirt FoldingFrom the PDF
Figure 21 Final states for Shirt Folding
Figure 21 Final states for Shirt FoldingFrom the PDF

Verification

Claims the paper makes

12 Valid · 4 Not valid · 14 Not checked · 115 Own results and interpretations

79 more claims · 3 with their cited sources checked

Create a free account to see all 145 claims and to verify the 5 references still open.

See all claims — free account
  • It may state two separate claims as one. 39
  • Probably not a substantial claim. 33
  • The quote only partly states this claim. 4

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 40 references

Research works printed but not linked: 4

  1. Janner M Du Y Tenenbaum J et al. (2022b) Planning with diffusion for flexible behavior synthesis. In: Chaudhuri K Jegelka S Song L et al. (eds) In: Proceedings of the 39th international conference on machine learning proceedings of machine learning research Baltimore MD 17–23 July 2022. PMLR.
  2. 10.1007/978-3-540-30301-5 · doi:10.1007/978-3-540-30301-5
  3. Wang Z Hunt JJ Zhou M (2023) Diffusion policies as an expressive policy class for offline reinforcement learning. In: The eleventh international conference on learning representations Vienna Austria 1 May 2023. URL: https://openreview.net/forum?id=AHvFDPi-FA.
  4. Welling M Teh YW (2011) Bayesian learning via stochastic gradient Langevin dynamics. In: Proceedings of the 28th international conference on machine learning (ICML-11) Bellevue WA 28 June–2 July 2011 pp. 681–688.

Cited by

See all 613

Metadata

Authors
Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., & Song, S.
Affiliations
  • Columbia University (US)
  • Toyota Research Institute (US)
  • Massachusetts Institute of Technology (US)
Corresponding author
Chi, Cheng
Venue
The International Journal of Robotics Research
Published
2024-10-11
Volume
44
Issue
10-11
Pages
1684-1704
ISSN
  • 0278-3649 (ISSN-L)
  • 1741-3176
OpenAlex
W4403337227
Cited by
613
References
40
Citations by year
2026: 335 · 2025: 259 · 2024: 17
FWCI
89.79
Type
journal article
Language
en
Open access
closed
Licence
journals.sagepub.com/page/policies/text-and-data-mining-license (tdm)
Locations
The International Journal of Robotics Research (publishedVersion)
Funders
  • Division of Civil, Mechanical and Manufacturing Innovation (2037101)
  • Toyota Research Institute
  • Division of Information and Intelligent Systems (2132519)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
  • Physical Sciences › Engineering › Control and Systems Engineering › Robot Manipulation and Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
diffusion policy, visuomotor policy learning, robotic manipulation, Langevin dynamics, visual conditioning, model predictive control, high-dimensional action spaces, physical robots

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Diffusion policy: Visuomotor policy learning via action diffusion. https://app.cochat.ai/papers/10.1177/02783649241273668
The paper
Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., & Song, S. (2025). Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research. https://doi.org/10.1177/02783649241273668

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.