Retrieve-Augmented Generation for Speeding up Diffusion Policy without Additional Training

Simon Holk; Sodtavilan Odonchimed; Tatsuya Matsushima; Yusuke Iwasawa; Yutaka Matsuo

REVIEW 3 major objections 7 minor 1 cited by

Retrieve-Augmented Generation for Speeding up Diffusion Policy without Additional Training

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A pretrained Diffusion Policy can be run up to 20 times faster by retrieving the nearest expert action from a vector database and starting denoising from it, without any additional training.

desk verdict Useful SDEdit-style acceleration for diffusion policies, but the VE update rule is suspect and missing baselines muddy where the gains come from. read the letter →

arxiv 2507.21452 v1 pith:2XBFKOGY submitted 2025-07-29 cs.LG cs.RO

Sodtavilan Odonchimed , Tatsuya Matsushima , Simon Holk , Yusuke Iwasawa , Yutaka Matsuo This is my paper

classification cs.LGcs.RO

keywords diffusionpolicyretrieval-augmentedgenerationinferenceaccelerationimitationlearningconsistencyscore-basedgenerativemodelsrobotmanipulation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes RAGDP, a way to make pretrained Diffusion Policy robot controllers generate actions faster without any additional training. The idea is to store the expert demonstrations used to train the policy as a vector database, look up the demonstration closest to the current observation at each inference, and start the diffusion denoising process from that retrieved action instead of from pure noise. Because the starting point is already near a good action, far fewer denoising steps are needed; only a fraction chosen by a hyperparameter r of the original steps is run. The authors report that on simulated manipulation benchmarks with state observations, this preserves accuracy at high speedups: at 20x speedup the VP variant keeps 64% of the original task completion rate where the efficient DPM++ sampler keeps only 17%, and the VE variant keeps 81.86% where the distillation-based Consistency Policy keeps 69.44%. The intended payoff is that existing diffusion policies can be adapted to real-time control simply by packaging their training data as a database, with no distillation cost.

What carries the argument

The load-bearing object is the retrieval database of observation-action pairs, built with the policy's own observation encoder (a normalization function for state observations), and searched by L2 nearest neighbor at every inference step. The second piece is the leap ratio r, which fixes how much of the denoising trajectory is skipped: the original T steps are cut to (1-r)T. RAGDP-VP constructs the starting iterate as A_{t,τ*} = $\sqrt$(α_{τ*}) A_ret + $\sqrt$(1-α_{τ*}) ε with α_{τ*}^2 + σ_{τ*}^2 = 1, while RAGDP-VE forms A_{t,T} = A_ret + σ_max ε and iterates a score update. The retrieved point is what lets the shortened chain stay near good behavior; the reduced step count is what produces the speedup.

What would settle it

Run the RAGDP-VE algorithm on the same EDM checkpoints but with the retrieved trajectory replaced by one drawn at random from the database, keeping all other settings fixed; if the reported recovery rates stay at 81%, retrieval is not what carries the gain. Also compare the update in Algorithm 2 with the standard Euler step for the probability-flow ODE; if the two produce different trajectories, the VE results rely on a solver choice that the paper does not justify.

Watch

Extended reading notes

Core claim

The paper's central claim is that the expensive part of a Diffusion Policy—generating an action by denoising Gaussian noise over roughly 100 steps—can be largely skipped without retraining, provided the policy has access to its own training demonstrations at inference time. RAGDP precomputes normalized embeddings of observation chunks and stores the corresponding action trajectories; at each control step it retrieves the trajectory whose observation is nearest in Euclidean distance, adds noise scaled to a chosen starting diffusion level, and runs only (1-r)T denoising steps. Two instantiations are given: RAGDP-VP for variance-preserving models, which mixes the retrieved action with noise according to the scheduler's alpha ratio, and RAGDP-VE for variance-exploding models, which adds maximum noise and then applies the learned score update. In the reported experiments the retrieved starting point buys enough accuracy that RAGDP beats both the fast DPM++ solver and the distilled Consistency Policy on the accuracy-speed trade-off, especially at high speedups.

Load-bearing premise

The method assumes the closest stored demonstration to the current observation is close enough to the correct action that a shortened denoising loop can refine it into a good trajectory, and it assumes the VE update in the second algorithm is a valid discretization of the score-based ODE.

Editorial extensions

If this is right

If RAGDP works as reported, any pretrained DDPM- or EDM-based Diffusion Policy can be accelerated 4 to 20 times simply by constructing a vector database from its training data, with no gradient updates.
At 20x speedup, RAGDP-VP on a DDPM model reports 64% recovery versus 17% for DPM++, so the method makes high-speed real-time deployment more viable without retraining.
RAGDP composes with existing solvers: combining RAGDP with DPM++ improves on DPM++ alone, so it is not an alternative to fast sampling but an add-on.
RAGDP-VE on EDM reports 81.86% recovery at 20x, compared to 69.44% for Consistency Policy, which suggests retrieval can match or beat distillation on state-observation control tasks without the training cost.
Because the method is agnostic to the policy's internal architecture, it can in principle also be applied on top of distilled consistency models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: The evidence only covers state observations, so the retrieval mechanism's reliability on image observations or out-of-distribution states is untested; if the embedding used for lookup is not locally smooth, the reported gains may shrink outside the demonstrated regime.
Editorial inference: The paper does not ablate retrieval quality directly; replacing retrieved actions with random database entries while keeping the step schedule would separate the benefit of a good starting point from the benefit of fewer steps.
Editorial inference: A natural extension is to set the leap ratio per query, using small r when the nearest neighbor is far and large r when it is close, which would let the method adapt the speed-accuracy trade-off dynamically rather than fixing it by hand.
Editorial inference: One practical boundary the paper explicitly acknowledges is that the database is built from training demonstrations, so deployment requires those demonstrations to be available; for robot policies shipped without their data, the method cannot be applied.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's the one-line: RAGDP is an SDEdit-style trick for diffusion policies—retrieve the nearest training action, add noise, and denoise for fewer steps. On robomimic it shows real accuracy–speed gains over DPM++ and Consistency Policy. But the VE variant has a suspicious update rule that needs correcting before its numbers are trustworthy.

What's new: the application of retrieval-initialized partial denoising to pretrained Diffusion Policies, with both VP and VE handling. That gives a training-free acceleration option for robot learning. The experiments cover four robomimic tasks and the paper measures recovery rate, which is the right metric. The implementation details are clear enough to reproduce in principle, though code is not released.

The soft spots are real. First, Algorithm 2's update, A_{τ−Δτ} = A_τ + (σ_τ² − σ_{τ−Δτ}²) s_θ(...), is not the standard Euler step for the VE-SDE probability flow. The standard coefficient is σ_τ(σ_τ − σ_{τ−Δτ}), about half as large. Worse, the paper says the EDM model outputs denoised samples, not scores, so the s_θ in Algorithm 2 doesn't match the model's prediction without a scaling factor. If that's a misprint, fine; if not, the RAGDP-VE results—including the headline 81.86% recovery at ×20—are unsupported. Second, there are no baselines with random noise initialization (plain SDEdit) or retrieval-only k-NN. Without those, you can't tell whether the gains come from retrieval or from starting the denoising closer to the demonstration distribution. At high r, the output is close to the retrieved action; that point is acknowledged but not quantified. Third, no error bars or significance tests on the numbers, which matters for the small differences at ×4.

The citation pattern looks fine; the paper is not hiding its relation to SDEdit and R2-Diff.

Who this is for: anyone working on fast inference for diffusion policies. It deserves a serious referee—the idea is of practical interest and the VP results are plausible. I would not accept it as-is. The VE solver needs a derivation or correction, the baselines need to be added, and the claims need to be scaled back until that's done.

Referee Report

3 major / 7 minor

Summary. The paper proposes RAGDP (Retrieve-Augmented Generation for Diffusion Policies), a training-free inference-time method for accelerating pretrained diffusion policies. During a preprocessing stage, observation-action pairs are embedded and stored in a FAISS vector database; at inference, the current observation is embedded, the nearest expert action is retrieved, and that action is used as the starting point for a reduced number of denoising steps. Two variants are described: RAGDP-VP for VP-SDE/DDPM models and RAGDP-VE for VE-SDE/EDM models. Experiments on four robomimic tasks compare RAGDP against DPM++ and Consistency Policy, reporting recovery rates at 4x, 10x, and 20x speedups and claiming improved accuracy-speed trade-offs without additional training.

Significance. If the results hold, RAGDP is a simple and practical idea: it wraps any pretrained diffusion policy with a vector database and requires no training or distillation, making it orthogonal to existing fast samplers and consistency-distillation methods. The use of FAISS for retrieval is sensible, and the paper explicitly examines robustness to mixed-quality demonstration data. The main weaknesses are that the VE-SDE update in Algorithm 2 appears mathematically invalid as written, the experiments lack the ablations needed to attribute gains to retrieval, and all results are point estimates without error bars or significance tests. These issues currently limit the strength of the central claim.

major comments (3)

[Algorithm 2 / Sec. IV.A] The VE-SDE update in Algorithm 2, line 12 is not a valid discretization of the probability-flow ODE. For the VE-SDE, dx/dsigma = -sigma * s_theta(x,sigma), so the Euler step from sigma_tau to sigma_{tau-Delta_tau} should be A_{tau-Delta_tau} = A_tau + sigma_tau * (sigma_tau - sigma_{tau-Delta_tau}) * s_theta(A_tau, sigma_tau, O_t). The algorithm instead uses (sigma_tau^2 - sigma_{tau-Delta_tau}^2) * s_theta(...) = (sigma_tau - sigma_{tau-Delta_tau}) * (sigma_tau + sigma_{tau-Delta_tau}) * s_theta(...), which is roughly twice the correct coefficient for fine steps and increasingly inaccurate on the coarse two-to-four-step schedules used at 10x and 20x speedups. Moreover, Sec. IV.A states that the EDM model outputs denoised samples directly, whereas Eq. (5) defines s_theta as the score; if the network output is a denoised sample, the quantity called s_theta in Algorithm 2 must be rescaled by 1/sigma_tau^2, and no such scaling is shown. As written, the RAGDP-VE recovery rates in Tables III and IV, including the headline 81.86% versus 69.44% comparison with CP, rest on a solver that is either misprinted or mathematically invalid. The authors should correct the update, provide its derivation, and confirm that the reported results are produced by the corrected algorithm.
[Sec. III.B / Sec. IV.C] The experimental design does not include the two ablations needed to attribute the observed gains to retrieval. Since r=1 would execute the retrieved action unchanged, a k-NN-only baseline (r=1, no diffusion) should be reported; without it, the contribution of the denoising part is unclear. Similarly, an SDEdit-style baseline that injects the same noise level into a non-retrieved or random action, or a random-Gaussian initialization with the same reduced step count, is needed to show that the nearest-expert initialization rather than merely a good in-distribution starting point is responsible for the improvements. These baselines are load-bearing for the central claim that RAGDP improves the accuracy-speed trade-off over existing samplers.
[Sec. IV.B / Sec. IV.C] All reported results are point estimates without error bars, confidence intervals, or significance tests. For example, Table III's 20x comparison (65.32% vs. 57.38%) and Table IV's 20x recovery comparison (81.86% vs. 69.44%) could fall within run-to-run variability across the three training seeds. The paper should report standard deviations or confidence intervals across seeds and, where feasible, a paired test over the evaluation seeds, especially for the CP versus RAGDP-VE comparisons that support the main conclusions.

minor comments (7)

[Sec. III.B.1] The image-observation embedding variant is described but never evaluated; all robomimic experiments use state observations (Sec. IV.A), so statements about applicability to image-based Diffusion Policies go beyond the current evidence.
[Algorithm 1 / Algorithm 2] The variable T is used both for total episode steps and total denoising steps, which is confusing; please use distinct symbols for these two quantities.
[Algorithm 1, line 11] The DDPM update uses alpha_tau without overbars; clarify whether alpha_tau denotes the cumulative product or the per-step schedule, since the formula is otherwise ambiguous.
[Sec. IV.E] The text says RAGDP-VP and RAGDP-VE are compared 'with the DPM++ sampler for both trained EDM models,' but Table IV reports RAGDP-VP on a DDPM base and RAGDP-VE on an EDM base; please reconcile this inconsistency.
[Sec. IV.A / Table I] The relationship between CTM and CP is unclear: Table I labels the consistency model as CTM, while the text refers to Consistency Policy (CP). Clarify whether the evaluated method is the Consistency Policy distilled from an EDM teacher and cite consistently.
[Abstract / Sec. IV.D] The abstract says a 7% increase over distillation models, while Table III shows about 8 percentage points and Table IV shows about 12 percentage points at 20x; report whether these are absolute or relative differences and keep the numbers consistent.
[Throughout] Typographical errors such as 'immitation,' 'denosing,' and 'V AE' should be corrected, and the grammar of several sentences (e.g., the first sentence of Sec. II.A) should be polished.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RAGDP is a transparent empirical retrieval-initialization method whose gains are measured against external samplers and whose r-to-1 limit (executing the retrieved expert action) is explicitly disclosed in the paper.

full rationale

The paper's derivation chain contains no step in which a predicted quantity is defined in terms of its own inputs, nor any fitted parameter relabeled as a prediction. RAGDP-VP initializes the chain at A_{t,τ*} = √α_{τ*} A_ret + √(1−α_{τ*}) ε (Algorithm 1, line 8), and RAGDP-VE initializes at A_{t,T} = A_ret + σ_max ε (Algorithm 2, line 9); the retrieved action A_ret is an input to the algorithm, and the paper explicitly discloses the r→1 limit ('when r=1, the retrieved trajectory is executed as is'). The headline accuracy-speed claims are comparative empirical results against DPM++ and CP (Tables III and IV) with held-out evaluation seeds, so they do not reduce to a self-definition: the improvement over DPM++ at ×20 is an external, falsifiable benchmark result, not a quantity forced by construction. There are no load-bearing self-citations: all cited prior results (SDE formulation [16], EDM design [38], SDEdit initialization [39], DPM++ [9]) are external to this article, and none is invoked to forbid alternatives or to justify the core retrieval step. The paper itself cites SDEdit as the origin of the noise-and-denoise initialization idea, so the method is not a renamed prior result presented without provenance. The main apparent defect flagged in review is Algorithm 2's VE-SDE update (line 12), which uses (σ_τ² − σ_{τ−Δτ}²) s_θ rather than the standard Euler coefficient σ_τ(σ_τ − σ_{τ−Δτ}) for the probability-flow ODE, and whose s_θ is not reconciled with the EDM 'sample' prediction type stated in Section IV.A; however, an incorrect discretization (if real) is a correctness risk, not circular reasoning, and outside this pass's scope. Likewise, the observation that at large r the output is dominated by the retrieved training action is a transparent property of the method, and the absence of an 'execute-retrieved-action-only' baseline is a missing-baseline concern, not a circular construction. Accordingly, the paper is self-contained against external benchmarks and no circular step is exhibited.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central results rest on two domain assumptions: retrieval relevance in the observation embedding, and the validity of partial denoising from a retrieved action. The only free parameter is the leap ratio r, which is selected by hand per speedup. The VE-SDE update in Algorithm 2 is an ad hoc assumption because it is not derived or referenced.

free parameters (1)

leap ratio r = 0.75, 0.875, 0.95 for x4/x10/x20 speedups on EDM
Hand-chosen hyperparameter controlling how many denoising steps are skipped; the paper's main results are reported across r values corresponding to x4, x10, x20 speedups, and the conclusion that RAGDP maintains accuracy depends on selecting r appropriately. No automatic selection method is provided.

assumptions (3)

ad hoc to paper The VE-SDE update in Algorithm 2 is a valid discretization of the EDM sampler.
Algorithm 2 states A_{tau-Delta_tau} = A_tau + (sigma_tau^2 - sigma_{tau-Delta_tau}^2) s_theta(...) without derivation or citation; it differs from the standard Euler step for the probability-flow ODE, so sampling correctness is assumed.
domain assumption Nearest-neighbor retrieval in raw normalized state space returns expert actions that are useful initializations for partial denoising in the tested tasks.
The entire method depends on the retrieved expert action having high task relevance; the paper only evaluates in-distribution robomimic state-observation tasks and does not measure retrieval distances or out-of-distribution behavior.
domain assumption The score function of the pretrained diffusion policy remains valid for inputs that mix a retrieved action with Gaussian noise at intermediate noise levels.
RAGDP initializes at tau_0 or sigma_max from a training-database action, which is not a typical pure-noise input seen during DP training; the paper relies on SDEdit-like behavior without a dedicated analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Retrieve-Augmented Generation for Speeding up Diffusion Policy without Additional Training." pith.science (2026). https://pith.science/paper/2XBFKOGY

@misc{pith2026250721452,
  author       = {Pith},
  title        = {Pith review of: Retrieve-Augmented Generation for Speeding up Diffusion Policy without Additional Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XBFKOGY}},
  note         = {Machine review of arXiv:2507.21452}
}

read the original abstract

Diffusion Policies (DPs) have attracted attention for their ability to achieve significant accuracy improvements in various imitation learning tasks. However, DPs depend on Diffusion Models, which require multiple noise removal steps to generate a single action, resulting in long generation times. To solve this problem, knowledge distillation-based methods such as Consistency Policy (CP) have been proposed. However, these methods require a significant amount of training time, especially for difficult tasks. In this study, we propose RAGDP (Retrieve-Augmented Generation for Diffusion Policies) as a novel framework that eliminates the need for additional training using a knowledge base to expedite the inference of pre-trained DPs. In concrete, RAGDP encodes observation-action pairs through the DP encoder to construct a vector database of expert demonstrations. During inference, the current observation is embedded, and the most similar expert action is extracted. This extracted action is combined with an intermediate noise removal step to reduce the number of steps required compared to the original diffusion step. We show that by using RAGDP with the base model and existing acceleration methods, we improve the accuracy and speed trade-off with no additional training. Even when accelerating the models 20 times, RAGDP maintains an advantage in accuracy, with a 7% increase over distillation models such as CP.

Figures

Figures reproduced from arXiv: 2507.21452 by the authors.

**Figure 1.** Diffusion-based Policies and RAGDP The Diffusion Models generate samples by removing noise from σmax to σmin over T steps. There are two types of noise removal methods: VP-SDE and VE-SDE. RAGDP adapts to each Diffusion Model and can generate actions using two methods (RAGDP-VP and RAGDP-VE). The proposed method can be accelerated by obtaining neighboring values from the knowledge base and reducing the number of nois… view at source ↗

**Figure 2.** a) Diffusion Policy Representation: Diffusion Policy behaves as a diffusion model that takes data observed from the environment as conditional input and outputs trajectory data. It generates samples by removing noise from σmax to σmin over T steps. b) Observation and Prediction Horizons: The conditional input is Ot, chunked by To steps of observed data ot, and the generated behavior is At, chunked by Tp steps of act… view at source ↗

**Figure 3.** The robustness of RAGDP: The graph displays the robustness when smaller sampling steps are used. We investigate using RAGDP with different base models and samplers. From the top, DDPM model using a DDPM sampler, DDPM model using a DPM++ sampler, and EDM model using a DPM++ sampler. The x-axis of the graph represents the number of sampling steps; the smaller the number, the faster the speed. y-axis represents the acc… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Inference speed and accuracy: 3 seeds average rewards are shown as accuracy on the vertical axis and sampling speed on the horizontal axis. Speeds were measured on RTX 3060. The upper left direction of the graph indicates better performance. The results show a comparis…

**Figure 6.** Figure 6: Accuracy and hyperparameter r: The effect of hyperparameter r on accuracy of the EDM model is shown. It can be seen that the larger the r, the faster the generation speed increases, but the accuracy tends to decrease. a greater extent. Calculating the average recovery …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Difference-Aware Retrieval Policies for Imitation Learning
cs.RO 2026-06 conditional novelty 6.0 of 10

DARP reparameterizes imitation learning around local neighborhood structure using k-NN expert states, actions, and relative distance vectors, delivering 15-46% gains over behavior cloning in control and manipulation tasks.

Reference graph

Works this paper leans on

45 extracted references · 20 canonical work pages · cited by 1 Pith paper

[1]

Is imitation learning the route to humanoid robots?

S. Schaal, “Is imitation learning the route to humanoid robots?” Trends in Cognitive Sciences, vol. 3, no. 6, pp. 233–242, 1999

work page 1999
[2]

An algorithmic perspective on imitation learning,

T. Osa, J. Pajarinen, G. Neumann, J. A. Bagnell, P. Abbeel, and J. Peters, “An algorithmic perspective on imitation learning,”Foun- dations and Trends in Robotics, vol. 7, no. 1–2, 1–179, 2018,ISSN: 1935-8261

work page 2018
[3]

O. M. Team et al.,Octo: An open-source generalist robot policy,

work page
[4]

Diffusion policy: Visuomotor policy learning via action diffusion,

C. Chi et al., “Diffusion policy: Visuomotor policy learning via action diffusion,”ArXiv, vol. abs/2303.04137, 2023

arXiv 2023
[5]

Chen, H.-C

S.-F. Chen, H.-C. Wang, M.-H. Hsu, C.-M. Lai, and S.-H. Sun, Diffusion model-augmented behavioral cloning, 2024. arXiv:2302. 13335 [cs.LG]

work page 2024
[6]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”ArXiv, vol. abs/2006.11239, 2020

arXiv 2006
[8]

J. Song, C. Meng, and S. Ermon,Denoising diffusion implicit models,

work page
[9]

C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu,Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models,

work page

Show all 45 references

[10]

Y . Song, P. Dhariwal, M. Chen, and I. Sutskever,Consistency models,
[11]

Kim et al.,Consistency trajectory models: Learning probabil- ity flow ode trajectory of diffusion, 2024

D. Kim et al.,Consistency trajectory models: Learning probabil- ity flow ode trajectory of diffusion, 2024. arXiv:2310 . 02279 [cs.LG]

2024
[12]

Salimans and J

T. Salimans and J. Ho,Progressive distillation for fast sampling of diffusion models, 2022. arXiv:2202.00512 [cs.LG]

2022 arXiv
[13]

A reduction of imitation learning and structured prediction to no-regret online learning,

S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proceedings of the fourteenth international conference on artificial intelligence and statistics, JMLR Workshop and Conference Proceed- ings, 2011...

2011
[14]

arXiv:2303.01469 [cs.LG]

arXiv
[15]

Lewis et al.,Retrieval-augmented generation for knowledge- intensive nlp tasks, 2021

P. Lewis et al.,Retrieval-augmented generation for knowledge- intensive nlp tasks, 2021. arXiv:2005.11401 [cs.CL]

2021 arXiv
[16]

Score-based generative modeling through stochastic differential equations,

Y . Song, J. N. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,”ArXiv, vol. abs/2011.13456, 2020

2011 arXiv
[17]

High-resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 10 674–10 685

2022
[18]

Toward the fundamental limits of imitation learning,

N. Rajaraman, L. Yang, J. Jiao, and K. Ramchandran, “Toward the fundamental limits of imitation learning,”Advances in Neural Information Processing Systems, vol. 33, pp. 2914–2924, 2020

2020
[19]

X. Li, C. Gu, S. Xie, Y . Bai, W. Zhang, and Z. Wang,Tuning-free visual customization via view iterative self-attention control, 2024. arXiv:2406.06258 [cs.CV]

2024 arXiv
[20]

Janner, Y

M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine,Planning with diffusion for flexible behavior synthesis, 2022. arXiv:2205.09991 [cs.LG]

2022 arXiv
[21]

Torabi, G

F. Torabi, G. Warnell, and P. Stone,Behavioral cloning from obser- vation, 2018. arXiv:1805.01954 [cs.AI]

2018 arXiv
[22]

J. Ho, C. Saharia, W. Chan, D. J. Fleet, M. Norouzi, and T. Salimans, Cascaded diffusion models for high fidelity image generation, 2021. arXiv:2106.15282 [cs.CV]

2021 arXiv
[23]

N. M. M. Shafiullah, Z. J. Cui, A. Altanzaya, and L. Pinto,Behavior transformers: Cloningkmodes with one stone, 2022. arXiv:2206. 11251 [cs.LG]

2022
[24]

Prasad, K

A. Prasad, K. Lin, J. Wu, L. Zhou, and J. Bohg,Consistency policy: Accelerated visuomotor policies via consistency distillation, 2024. arXiv:2405.07503 [cs.RO]

2024 arXiv
[25]

Wang et al.,One-step diffusion policy: Fast visuomotor policies via diffusion distillation, 2024

Z. Wang et al.,One-step diffusion policy: Fast visuomotor policies via diffusion distillation, 2024. arXiv:2410.21257 [cs.RO]

2024 arXiv
[26]

Florence et al.,Implicit behavioral cloning, 2021

P. Florence et al.,Implicit behavioral cloning, 2021. arXiv:2109. 00137 [cs.RO]

2021
[27]

C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu,Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022. arXiv:2206.00927 [cs.LG]

2022 arXiv
[28]

Rombach, A

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, High-resolution image synthesis with latent diffusion models, 2022. arXiv:2112.10752 [cs.CV]

2022 arXiv
[29]

van den Oord, O

A. van den Oord, O. Vinyals, and K. Kavukcuoglu,Neural discrete representation learning, 2018. arXiv:1711.00937 [cs.LG]

2018 arXiv
[30]

Diffusion models: A comprehensive survey of methods and applications,

L. Yang et al., “Diffusion models: A comprehensive survey of methods and applications,”ACM Computing Surveys, vol. 56, pp. 1 –39, 2022

2022
[31]

Nasiriany, T

S. Nasiriany, T. Gao, A. Mandlekar, and Y . Zhu,Learning and retrieval from prior data for skill-based imitation learning, 2022. arXiv:2210.11435 [cs.LG]

2022 arXiv
[32]

M. Du, S. Nair, D. Sadigh, and C. Finn,Behavior retrieval: Few- shot imitation learning by querying unlabeled datasets, 2023. arXiv: 2304.08742 [cs.RO]

2023 arXiv
[33]

Hejna et al.,Robot data curation with mutual information estima- tors, 2025

J. Hejna et al.,Robot data curation with mutual information estima- tors, 2025. arXiv:2502.08623 [cs.RO]

2025 arXiv
[34]

A. Xie, O. Rybkin, D. Sadigh, and C. Finn,Latent diffusion planning for imitation learning, 2025. arXiv:2504.16925 [cs.RO]

2025 arXiv
[35]

Read: Retrieval-enhanced asymmetric diffusion for motion planning,

T. Oba, M. Walter, and N. Ukita, “Read: Retrieval-enhanced asymmetric diffusion for motion planning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 17 974–17 984

2024
[36]

Zou et al.,Parco: Part-coordinating text-to-motion synthesis,

Q. Zou et al.,Parco: Part-coordinating text-to-motion synthesis,
[37]

Zhang et al.,Motiondiffuse: Text-driven human motion generation with diffusion model, 2022

M. Zhang et al.,Motiondiffuse: Text-driven human motion generation with diffusion model, 2022. arXiv:2208.15001 [cs.CV]

2022 arXiv
[38]

Oba and N

T. Oba and N. Ukita,R2-diff: Denoising by diffusion as a refinement of retrieved motion for image-based motion prediction, 2023. arXiv: 2306.09483 [cs.CV]

2023 arXiv
[39]

Meng et al.,Sdedit: Guided image synthesis and editing with stochastic differential equations, 2022

C. Meng et al.,Sdedit: Guided image synthesis and editing with stochastic differential equations, 2022. arXiv:2108 . 01073 [cs.CV]

2022
[40]

Johnson, M

J. Johnson, M. Douze, and H. J ´egou,Billion-scale similarity search with gpus, 2017. arXiv:1702.08734 [cs.CV]

2017 arXiv
[41]

arXiv:2403.18512 [cs.CV]

arXiv
[43]

Elucidating the design space of diffusion-based generative models,

T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,”ArXiv, vol. abs/2206.00364, 2022

2022 arXiv
[46]

What matters in learning from offline human demonstrations for robot manipulation,

A. Mandlekar et al., “What matters in learning from offline human demonstrations for robot manipulation,” inConference on Robot Learning, 2021

2021
[2022]

arXiv:2010.02502 [cs.LG]

2010 arXiv
[2023]

arXiv:2211.01095 [cs.LG]

arXiv
[2024]

arXiv:2405.12213 [cs.RO]

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Is imitation learning the route to humanoid robots?

S. Schaal, “Is imitation learning the route to humanoid robots?” Trends in Cognitive Sciences, vol. 3, no. 6, pp. 233–242, 1999

work page 1999

[2] [2]

An algorithmic perspective on imitation learning,

T. Osa, J. Pajarinen, G. Neumann, J. A. Bagnell, P. Abbeel, and J. Peters, “An algorithmic perspective on imitation learning,”Foun- dations and Trends in Robotics, vol. 7, no. 1–2, 1–179, 2018,ISSN: 1935-8261

work page 2018

[3] [3]

O. M. Team et al.,Octo: An open-source generalist robot policy,

work page

[4] [4]

Diffusion policy: Visuomotor policy learning via action diffusion,

C. Chi et al., “Diffusion policy: Visuomotor policy learning via action diffusion,”ArXiv, vol. abs/2303.04137, 2023

arXiv 2023

[5] [5]

Chen, H.-C

S.-F. Chen, H.-C. Wang, M.-H. Hsu, C.-M. Lai, and S.-H. Sun, Diffusion model-augmented behavioral cloning, 2024. arXiv:2302. 13335 [cs.LG]

work page 2024

[6] [6]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”ArXiv, vol. abs/2006.11239, 2020

arXiv 2006

[7] [8]

J. Song, C. Meng, and S. Ermon,Denoising diffusion implicit models,

work page

[8] [9]

C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu,Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models,

work page