Diffusion Policy

Diffusion Policydiffusion policyA diffusion policy is a policy generating a whole action sequence by conditional denoising diffusion instead of regressing one action. Introduced by Chi et al. in 2023, it represents multimodal demonstrations without averaging their modes. It executes only the first few steps of each sequence, then observes again and replans. is a visuomotor robot policy that generates an action sequence by conditional denoising, from Chi et al. (2023). Its extended journal version evaluates 15 tasks and reports a 46.9% average relative improvement in simulation, using the best baseline and backbone per column, multi-human data excluded^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023.

Paper figure: Tony Z. Zhao et al. / arXiv. Licence: CC BY 4.0.

Receding-horizon control

Diffusion Policy executes $T_a$ actions of a $T_p$-action prediction, then observes and replans. The CNN configuration executes 8 of 16 on many tasks, and the next inference can warm-start from the previous prediction^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023.

$T_a$ acts like ACT's chunk size: large values give smooth motion that reacts slowly, and small values react fast but resample often, so each sample can switch mode^{Source 4}4 Zhao et al. 2023, Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. Tony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023. At the policy's 10 Hz command rate, $T_a = 8$ commits the robot for 0.8 seconds^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023. Try “Never rethink”: the dashed tails vanish.

Multimodal demonstrations

When one operator passes an obstacle on the left and another on the right, mean-squared-error regression averages the demonstrations into a collision. Energy-based implicit policies often beat such explicit policies^{Source 10}10 Florence et al. 2021, Implicit Behavioral Cloning. Pete Florence, Corey Lynch, Andy Zeng et al., 2021, and diffusion models learn an expressive distribution over the joint action space^{Source 11}11 Pearce et al. 2023, Imitating Human Behaviour with Diffusion Models. Tim Pearce, Tabish Rashid, Anssi Kanervisto et al., ICLR 2023. Pan et al. (2025) contest this motivation: generative policies owe their gains to supervised iterative computation, and a two-step regression policy essentially matches flow policies^{Source 12}12 Pan et al. 2025, Much Ado About Noising: Dispelling the Myths of Generative Robotic Control. Chaoyi Pan, Giri Anantharaman, Nai-Chieh Huang et al., 2025. Fine-tuned OpenVLA-OFT, an L1-regression VLA, beat from-scratch Diffusion Policy and ACT by up to 15 points of average success^{Source 13}13 Kim et al. 2025, Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. Moo Jin Kim, Chelsea Finn, Percy Liang, 2025.

Denoising in action space

The policy is a denoising diffusion probabilistic model (DDPM)^{Source 15}15 Ho et al. 2020, Denoising Diffusion Probabilistic Models. Jonathan Ho, Ajay Jain, Pieter Abbeel, 2020 conditioned on recent observations. Where a regression head computes $a_t = f_\theta(o_t)$, Diffusion Policy models $p(A_t \mid O_t)$ over the next $T_p$ actions. Inference starts from Gaussian noise and repeatedly subtracts a fraction of the noise that the denoiser $\epsilon_\theta$ predicts,

until the sample is a valid action chunk. Training uses noise-prediction MSE with improved DDPM's square-cosine schedule^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023 ^{Source 16}16 Nichol et al. 2021, Improved Denoising Diffusion Probabilistic Models. Alex Nichol, Prafulla Dhariwal, 2021. Real-world runs use DDIM, which keeps DDPM training but samples more efficiently^{Source 17}17 Song et al. 2021, Denoising Diffusion Implicit Models. Jiaming Song, Chenlin Meng, Stefano Ermon, ICLR 2021, with 10 inference iterations and 0.1-second latency on an Nvidia 3080 in the main text and 16 in the appendix^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023.

CNN and transformer variants

The CNN backbone adapts Diffuser's one-dimensional temporal U-Net over state-action trajectories^{Source 3}3 Janner et al. 2022, Planning with Diffusion for Flexible Behavior Synthesis. Michael Janner, Yilun Du, Joshua B. Tenenbaum et al., ICML 2022 to model only $p(A_t \mid O_t)$ and predict only actions^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023, injecting observations through FiLM's feature-wise affine transformation^{Source 18}18 Perez et al. 2018, FiLM: Visual Reasoning with a General Conditioning Layer. Ethan Perez, Florian Strub, Harm de Vries et al., AAAI 2018:

The CNN is the recommended default and works on most tasks without much tuning. The transformer gives most of the best state-based results, especially with fast action changes, but costs more tuning, and adding layers sometimes hurts^{Source 1}1 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Zhenjia Xu, Siyuan Feng et al., 2023. ScaleDP confirms the layer problem and scales the transformer from 10 million to 1 billion parameters^{Source 19}19 Zhu et al. 2024, Scaling Diffusion Policy in Transformer to 1 Billion Parameters for Robotic Manipulation. Minjie Zhu, Yichen Zhu, Jinming Li et al., 2024.

Variants and scale

DP3 conditions on sparse point clouds and handles most of 72 simulation tasks with 10 demonstrations^{Source 20}20 Ze et al. 2024, 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations. Yanjie Ze, Gu Zhang, Kangning Zhang et al., RSS 2024, and Equivariant Diffusion Policy builds in SO(2) symmetry for 21.9% higher average success^{Source 21}21 Wang et al. 2024, Equivariant Diffusion Policy. Dian Wang, Stephen Hart, David Surovik et al., CoRL 2024. ALOHA Unleashed pairs Diffusion Policies with large-scale ALOHA 2 data for deformable, contact-rich bimanual tasks^{Source 22}22 Zhao et al. 2024, ALOHA Unleashed: A Simple Recipe for Robot Dexterity. Tony Z. Zhao, Jonathan Tompson, Danny Driess et al., 2024, and TRI's Large Behavior Models scale the recipe to approximately 1,700 hours and match from-scratch policies with 3 to 5 times less data^{Source 23}23 TRI LBM Team et al. 2025, A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation. TRI LBM Team, Jose Barreiros, Andrew Beaulieu et al., 2025.

Paper figure: Yanjie Ze et al. / arXiv. Licence: CC BY 4.0.

Distillation

Iterative denoising is slow for mobile manipulators and quadrotors without high-end GPUs. Consistency models generate in one step by design^{Source 24}24 Song et al. 2023, Consistency Models. Yang Song, Prafulla Dhariwal, Mark Chen et al., ICML 2023, and Consistency Policy distills a Diffusion Policy teacher into one- or three-step generation; on Robomimic Square, three teachers at 0.84 to 0.92 success yielded students at 0.88 to 0.92^{Source 5}5 Prasad et al. 2024, Consistency Policy: Accelerated Visuomotor Policies via Consistency Distillation. Aaditya Prasad, Kevin Lin, Jimmy Wu et al., 2024.

One-Step Diffusion Policy lifts action prediction from 1.5 Hz to 62 Hz on real Franka tasks^{Source 6}6 Wang et al. 2024, One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation. Zhendong Wang, Zhaoshuo Li, Ajay Mandlekar et al., 2024. Drifting Models (Kaiming He and colleagues, February 2026) train a one-step generator without distillation; swapped into Diffusion Policy, one step matched or beat the 100-step original on 10 of 14 simulated task settings^{Source 30}30 Deng et al. 2026, Generative Modeling via Drifting. Mingyang Deng, He Li, Tianhong Li et al., 2026.

Flow matching

π0 replaced the DDPM with conditional flow matchingflow matchingFlow matching is a generative modeling recipe that learns a time-dependent vector field carrying noise to data, by regressing against the conditional flow instead of estimating a score. In pi0, which brought it to robot control, an action expert trained this way denoises continuous action chunks. Integrating the field as an ODE for a handful of steps is fast enough for high-frequency control and keeps diffusion policies' multimodality., which regresses vector fields of fixed conditional probability paths^{Source 26}26 Lipman et al. 2022, Flow Matching for Generative Modeling. Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu et al., 2022, on a straight path as in rectified flow^{Source 27}27 Liu et al. 2022, Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. Xingchao Liu, Chengyue Gong, Qiang Liu, 2022. It samples noise $\varepsilon$ and forms the interpolant

then trains the network to predict $u$, drawing $\tau$ from a beta distribution weighted toward the noisy end. Sampling integrates from $\tau = 0$ to $\tau = 1$ with 10 forward Euler steps^{Source 8}8 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Bilinear Flow Policy (Georgia Tech, University of Washington and Toyota Research Institute, October 5, 2026) targets goals outside the training distribution: across five simulated tasks it reached 2.63 times the out-of-distribution success of a goal-conditioned flow policy, and on two real tasks 50% against 18%^{Source 31}31 Jung et al. 2026, Bilinear Flow Policy: Distributional Extrapolation for Goal-Conditioned Visuomotor Imitation. Wonsuhk Jung, Sundhar Vinodh Sangeetha, Chen Xu et al., 2026.

Real-time chunking

Real-Time Chunkingreal-time chunkingReal-time chunking is an inference-time scheme keeping chunked policies responsive when inference is slower than control. While the robot finishes the current chunk, the policy conditions the next on the actions that will already have executed when inference returns, holding that prefix fixed. Black, Galliker and Levine introduced it for flow-matching policies; it tracks moving targets under delays that break naive chunking and temporal ensembling. generates the next chunk while the current one executes, holding average throughput at +100 and +200 ms of injected delay on six real π0.5 tasks^{Source 9}9 Black et al. 2025, Real-Time Execution of Action Chunking Flow Policies. Kevin Black, Manuel Y. Galliker, Sergey Levine, 2025, and training-time RTC removes its inference overhead^{Source 28}28 Black et al. 2025, Training-Time Action Conditioning for Efficient Real-Time Chunking. Kevin Black, Allen Z. Ren, Michael Equi et al., 2025.

Limitations

The original paper conditions on observations alone, on single- and multi-task benchmarks and separate UR5 and Franka setups^{Source 2}2 Chi et al. 2023, Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Cheng Chi, Siyuan Feng, Yilun Du et al., 2023; Octo adds language or goal-image conditioning^{Source 7}7 Octo Model Team et al. 2024, Octo: An Open-Source Generalist Robot Policy. Octo Model Team, Dibya Ghosh, Homer Walke et al., 2024, and π0 a pretrained vision-language backbone^{Source 8}8 Black et al. 2024, π0: A Vision-Language-Action Flow Model for General Robot Control. Kevin Black, Noah Brown, Danny Driess et al., RSS 2025, 2024.

Reactivity is bounded by $T_a$: events inside a committed segment wait for the next inference^{Source 14}14 Liu et al. 2024, Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. Yuejiang Liu, Jubayer Ibn Hamid, Annie Xie et al., 2024. Sampling cost remains, though one-step successors such as MP1 cut it to 6.8 ms^{Source 25}25 Sheng et al. 2026, MP1: MeanFlow Tames Policy Learning in 1-step for Robotic Manipulation. Juyi Sheng, Ziyi Wang, Peiming Li et al., AAAI 2026. With sparse data it acts like an action lookup table, returning training actions even for images of cats and dogs^{Source 29}29 He et al. 2025, Demystifying Diffusion Policies: Action Memorization and Simple Lookup Table Alternatives. Chengyang He, Xu Liu, Gadiel Sznaier Camps et al., 2025.