{"work":{"id":"2dce18e6-f07a-4f57-8a81-e71c3e6a293c","openalex_id":"https://openalex.org/W4323651335","doi":"10.48550/arxiv.2303.04137","arxiv_id":"2303.04137","raw_key":null,"title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion","authors":null,"authors_text":"Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel","year":2023,"venue":"cs.RO","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 12 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 is publicly available diffusion-policy.cs.columbia.edu","external_url":"https://arxiv.org/abs/2303.04137","cited_by_count":13,"metadata_source":"pith","metadata_fetched_at":"2026-08-05T02:28:24.338817+00:00","pith_arxiv_id":"2303.04137","created_at":"2026-05-09T06:10:42.430902+00:00","updated_at":"2026-08-05T02:28:24.338817+00:00","title_quality_ok":true,"display_title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion","render_title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"},"hub":{"state":{"work_id":"2dce18e6-f07a-4f57-8a81-e71c3e6a293c","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":104,"external_cited_by_count":13,"distinct_field_count":6,"first_pith_cited_at":"2023-05-22T17:57:41+00:00","last_pith_cited_at":"2026-07-09T09:24:30+00:00","author_build_status":"needed","summary_status":"needed","contexts_status":"needed","graph_status":"needed","ask_index_status":"needed","reader_status":"not_needed","recognition_status":"not_needed","updated_at":"2026-08-22T16:49:32.355990+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"background","n":10},{"context_role":"method","n":7},{"context_role":"baseline","n":1},{"context_role":"dataset","n":1},{"context_role":"extension","n":1}],"polarity_counts":[{"context_polarity":"background","n":9},{"context_polarity":"use_method","n":7},{"context_polarity":"baseline","n":1},{"context_polarity":"extend","n":1},{"context_polarity":"unclear","n":1},{"context_polarity":"use_dataset","n":1}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion","claims":[{"claim_text":"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 12 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 dynam","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"ceive, modulate, and record internal grasping force, external interaction wrench, and tactile interaction during demon- stration. In this sense, OmniUMI does not merely measure physical variables; it organizes their acquisition around a human-aligned interface for natural interaction regulation. At the learning and deployment level, we extend dif- fusion policy (DP) [6] to incorporate multimodal observa- tions, including visual, tactile, and force-related signals, en- abling policies to reason j","claim_type":"extension","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"9 53.9 42.0 32.0 2.89 TaKSIE [13] WACV, 2025 Plan 1 90.4 73.9 61.7 51.2 40.8 3.18 HULC++ [26] ICRA, 2023 Plan 1 93.0 79.0 64.0 52.0 40.0 3.30 RoboTron-Mani [40] ICCV, 2025 3D input 1 94.7 80.3 65.1 51.4 39.0 3.31 DaDu-Corki-SW [9] ISCA, 2025 - 1 92.3 80.0 67.4 56.6 45.8 3.42 RoboUniView (default) [23] arXiv, 2024 - 195.4 82.768.5 56.4 46.1 3.49 DTP [7] RA-L, 2025 Plan 1 92.4 81.970.2 60.3 50.9 3.55 NFE = 4 methods LaDi-WM [10] arXiv, 2025 WM≈4 92.7 83.1 72.1 61.2 54.1 3.63 LightDP [38] ICCV, 202","claim_type":"baseline","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"grounding [9, 290], world modeling [291, 292], and temporal prediction [293]. Since web data primarily consistsofvision-languagepairsandismainlyusedtoenhancefoundationmodelcapabilities, whichfallsoutside the scope of our survey, we focus primarily on human video in our discussion. Representative human video datasets include Ego4D [236], EPIC-KITCHENS-100 [277], and Something-Something V2 [275]. Although these datasets do not contain action labels directly usable for policy learning, they capture","claim_type":"dataset","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"network with data annotated at the sub-step level. Consequently, our method can complete com- plex tasks, like laundry folding, without requiring an external high-level policy, making the entire framework more end-to-end and demonstrating significant potential. Diffusion models. Diffusion models [48, 49, 50] have emerged as the dominant approach in visual generation. The Diffusion Policy [30] successfully applies the diffusion model to robot learning, demonstrating its ability to model multimoda","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"[2] primarily relied on autoregressive or behavior-cloning policies that predict actions step by step. More recently, dual-system VLA architectures [1, 9] have emerged, decoupling high-level multimodal reasoning [19, 22] from low-level action generation. In these systems, a perception-language back- bone [19] produces contextual representations, while a dedicated action expert is often implemented as a diffusion [23] or flow-matching head that predicts multiple future actions in parallel [24, 25","claim_type":"method","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"models in current usage; both are most usefully read here as design boundaries rather than additional roles. 10.1 Generative world simulators Generative world simulators form a boundary case. Systems inspired by large-scale video generation, interactive environment generation, and physical-AI simulation platforms are often called world models because they produce plausible future observations or playable environments [12, 49, 64], but photorealistic or interactive generation is not identical to ","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Diffusion Policy: Visuomotor Policy Learning via Action Diffusion because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (10 contexts).","role_counts":[{"n":10,"context_role":"background"},{"n":7,"context_role":"method"},{"n":1,"context_role":"baseline"},{"n":1,"context_role":"dataset"},{"n":1,"context_role":"extension"}]},"error":null,"updated_at":"2026-07-09T00:52:57.235735+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"31a9bd4d-7ff4-4abd-ae50-4d238cdcf285","orcid":null,"display_name":"Cheng Chi"},{"id":"5ab9dce0-e5a8-4982-8f67-a1d384e9986c","orcid":null,"display_name":"Zhenjia Xu"},{"id":"0d3d829c-ee2f-453a-bcd7-c5a168009f59","orcid":null,"display_name":"Siyuan Feng"},{"id":"f8532816-edc4-4aad-b739-881d427b2d7e","orcid":null,"display_name":"Eric Cousineau"},{"id":"b222acc7-ffce-4cd8-9940-9375e32ace22","orcid":null,"display_name":"Yilun Du"},{"id":"e154a17f-60d9-4108-b473-30086aaeb366","orcid":null,"display_name":"Benjamin Burchfiel"}]},"error":null,"updated_at":"2026-07-09T00:52:57.664901+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-07-09T00:53:07.083117+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","shared_citers":46},{"title":"OpenVLA: An Open-Source Vision-Language-Action Model","work_id":"3e7e65c5-5aed-4fe9-8414-2092bcb31cc7","shared_citers":43},{"title":"Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware","work_id":"6fe159e0-fa73-481a-88d4-4719c15140be","shared_citers":40},{"title":"RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control","work_id":"ff438a8a-8003-4fae-9131-acd418b3597b","shared_citers":34},{"title":"RT-1: Robotics Transformer for Real-World Control at Scale","work_id":"e11bda85-8531-46bc-a07f-d0ade3643ab1","shared_citers":31},{"title":"$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization","work_id":"d1ad7304-d09a-49bc-809e-846439f6aff9","shared_citers":29},{"title":"Octo: An Open-Source Generalist Robot Policy","work_id":"f9ca0722-8855-48c3-a27a-0eefb7e19253","shared_citers":28},{"title":"Open X-Embodiment: Robotic Learning Datasets and RT-X Models","work_id":"62f0fb6c-e6ae-4dc4-95a4-d9dd64b240e8","shared_citers":26},{"title":"Flow Matching for Generative Modeling","work_id":"6edb71c4-5d64-40af-a394-9757ea051a36","shared_citers":22},{"title":"Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success","work_id":"04f46bb3-4346-47e8-bf09-c75d91f96e87","shared_citers":18},{"title":"GR00T N1: An Open Foundation Model for Generalist Humanoid Robots","work_id":"e2db69c7-ee8a-4cb7-a761-7b8de1dfcf97","shared_citers":18},{"title":"Denoising Diffusion Implicit Models","work_id":"8fa2128b-d18c-405c-ac92-0e669cf89ac0","shared_citers":17},{"title":"FAST: Efficient Action Tokenization for Vision-Language-Action Models","work_id":"83a8f966-6cfa-4f21-81f3-87440aae238f","shared_citers":17},{"title":"LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning","work_id":"662203ad-084f-42c4-8e60-977b3173755b","shared_citers":16},{"title":"RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation","work_id":"12319725-bc7d-4c32-a229-ad270a7460bc","shared_citers":14},{"title":"SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics","work_id":"0c5e9314-5fa7-4613-ad12-605a71d561d2","shared_citers":14},{"title":"Do As I Can, Not As I Say: Grounding Language in Robotic Affordances","work_id":"037320f1-b0a9-4cbe-a639-bfb25409ce71","shared_citers":13},{"title":"World Action Models are Zero-shot Policies","work_id":"9a85fc69-74df-450e-94cd-69d186e9e830","shared_citers":13},{"title":"3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations","work_id":"bded01e1-c070-4537-a75a-ace4c75d0c95","shared_citers":12},{"title":"Decoupled Weight Decay Regularization","work_id":"07ef7360-d385-4033-83f7-8384a6325204","shared_citers":12},{"title":"Denoising Diffusion Probabilistic Models","work_id":"dc023f4e-7c79-471c-b713-deeb559ba010","shared_citers":12},{"title":"DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset","work_id":"13253de2-3d89-415c-8c2f-3adb25d4c337","shared_citers":12},{"title":"Planning with Diffusion for Flexible Behavior Synthesis","work_id":"38b2c635-b754-412a-a8f5-dfcf3e405c95","shared_citers":12},{"title":"What Matters in Learning from Offline Human Demonstrations for Robot Manipulation","work_id":"6a4c95c5-540e-4854-946d-c7c8a6c540ba","shared_citers":12}],"time_series":[{"n":5,"year":2023},{"n":8,"year":2024},{"n":7,"year":2025},{"n":83,"year":2026}],"dependency_candidates":[{"n":1,"role":"method","polarity":"use_method","paper_title":"Dynamic Execution Commitment of Vision-Language-Action Models","primary_cat":"cs.CV","context_text":"[2] primarily relied on autoregressive or behavior-cloning policies that predict actions step by step. More recently, dual-system VLA architectures [1, 9] have emerged, decoupling high-level multimodal reasoning [19, 22] from low-level action generation. In these systems, a perception-language back- bone [19] produces contextual representations, while a dedicated action expert is often implemented as a diffusion [23] or flow-matching head that predicts multiple future actions in parallel [24, 25, 26]. In practice, these systems often execute a predefined number of predicted actions before re-planning, with the execution horizon typically selected per task or benchmark [27, 28, 29]. For instance, π-0.5 [1] adopts different execution horizons across LIBERO subtasks [ 8] to balance performance and","citing_arxiv_id":"2605.11567"},{"n":1,"role":"baseline","polarity":"baseline","paper_title":"CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies","primary_cat":"cs.CV","context_text":"9 53.9 42.0 32.0 2.89 TaKSIE [13] WACV, 2025 Plan 1 90.4 73.9 61.7 51.2 40.8 3.18 HULC++ [26] ICRA, 2023 Plan 1 93.0 79.0 64.0 52.0 40.0 3.30 RoboTron-Mani [40] ICCV, 2025 3D input 1 94.7 80.3 65.1 51.4 39.0 3.31 DaDu-Corki-SW [9] ISCA, 2025 - 1 92.3 80.0 67.4 56.6 45.8 3.42 RoboUniView (default) [23] arXiv, 2024 - 195.4 82.768.5 56.4 46.1 3.49 DTP [7] RA-L, 2025 Plan 1 92.4 81.970.2 60.3 50.9 3.55 NFE = 4 methods LaDi-WM [10] arXiv, 2025 WM≈4 92.7 83.1 72.1 61.2 54.1 3.63 LightDP [38] ICCV, 2025 Distill 493.7 84.5 74.1 64.4 55.6 3.72 NFE≥10 methods RoboTron-Mani (DiT) [40] arXiv, 2025 3D input>1096.9 83.068.1 56.5 46.8 3.51 𝜋 ∗ 0.5 [11] CoRL, 2025 3B 10 90.4 78.1 67.961.2 54.73.52 MDT (default) [31] RSS, 2024 - 10 93.","citing_arxiv_id":"2604.24622"},{"n":1,"role":"method","polarity":"use_method","paper_title":"Mask World Model: Predicting What Matters for Robust Robot Policy Learning","primary_cat":"cs.RO","context_text":"for reproducibility. Timestep conditioning via AdaIN-style modulation.In- stead of standard adaptive normalization, MWM applies timestep-dependent scale and shift modulation after normal- ization(Huang & Belongie, 2017). 4 For a hidden activation x, we first normalize using RM- SNorm, then modulate: ¯x= RMSNorm(x), Modulate(¯x;s) = ¯x⊙ 1 +α(s) \u0001 +β(s) (7) Here, α(s) and β(s) are functions of the timestep embed- ding s, learned through a combination of a trainable scale- shift table and timestep projection. The operator ⊙ repre- sents element-wise multiplication. This design enhances stability when operating on normal- ized V AE latents by effectively handling their inherent vari- ability, while simultaneously preserving accurate timestep-","citing_arxiv_id":"2604.19683"},{"n":1,"role":"extension","polarity":"extend","paper_title":"OmniUMI: Towards Physically Grounded Robot Learning via Human-Aligned Multimodal Interaction","primary_cat":"cs.RO","context_text":"ceive, modulate, and record internal grasping force, external interaction wrench, and tactile interaction during demon- stration. In this sense, OmniUMI does not merely measure physical variables; it organizes their acquisition around a human-aligned interface for natural interaction regulation. At the learning and deployment level, we extend dif- fusion policy (DP) [6] to incorporate multimodal observa- tions, including visual, tactile, and force-related signals, en- abling policies to reason jointly over geometry and inter- action. On the deployment side, we translate policy out- puts into virtual targets and execute them via impedance- based control, enabling implicit and unified regulation of motion and force without explicit switching between con-","citing_arxiv_id":"2604.10647"},{"n":1,"role":"method","polarity":"use_method","paper_title":"Referring-Aware Visuomotor Policy Learning for Closed-Loop Manipulation","primary_cat":"cs.RO","context_text":"on the referring action P a: we simply force its transla- tion component atrans to coincide with the referring point P, while leaving the rotation component arot free to be optimized by the policy model. We then implement our trajectory-steeringstrategy through a masked-denoising process (Tseng et al., 2023; Kim et al., 2023),i.e., zt =M ⊙A known + (1− M)⊙z t (6) where zt denotes the intermediate noisy action vector at diffusion timestep t, ⊙ represents the Hadamard (element- wise) product, Aknown is a known action vector used to steer the denoising, and M is a binary mask indicating the indices to replace within the full noisy trajectory. As stated in Sec. 3.2, this strategy is applied in two stages: (i)","citing_arxiv_id":"2604.05544"},{"n":1,"role":"dataset","polarity":"use_dataset","paper_title":"A Survey on Vision-Language-Action Models: An Action Tokenization Perspective","primary_cat":"cs.RO","context_text":"grounding [9, 290], world modeling [291, 292], and temporal prediction [293]. Since web data primarily consistsofvision-languagepairsandismainlyusedtoenhancefoundationmodelcapabilities, whichfallsoutside the scope of our survey, we focus primarily on human video in our discussion. Representative human video datasets include Ego4D [236], EPIC-KITCHENS-100 [277], and Something-Something V2 [275]. Although these datasets do not contain action labels directly usable for policy learning, they capture diverse human-object interactions, complex manipulation skills, and rich physical common sense, which are valuable sources of world knowledge. Their scale and diversity enable pretraining of temporal visual encoders and facilitate the learning of action token representations.","citing_arxiv_id":"2507.01925"},{"n":1,"role":"method","polarity":"use_method","paper_title":"DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control","primary_cat":"cs.RO","context_text":"network with data annotated at the sub-step level. Consequently, our method can complete com- plex tasks, like laundry folding, without requiring an external high-level policy, making the entire framework more end-to-end and demonstrating significant potential. Diffusion models. Diffusion models [48, 49, 50] have emerged as the dominant approach in visual generation. The Diffusion Policy [30] successfully applies the diffusion model to robot learning, demonstrating its ability to model multimodal action distributions. Subsequent research has further developed the Diffusion Policy [51, 52, 53, 7, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63] by applying it to 3D environments [64, 65, 66, 67], scaling its capabilities [29], improving its efficiency [68, 53],","citing_arxiv_id":"2502.05855"},{"n":1,"role":"method","polarity":"use_method","paper_title":"RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots","primary_cat":"cs.RO","context_text":"control at scale. In arXiv preprint arXiv:2212.06817 , 2022. [3] Yevgen Chebotar, Quan Vuong, Karol Hausman, Fei Xia, Yao Lu, Alex Irpan, Aviral Kumar, Tianhe Yu, Alexander Herzog, Karl Pertsch, et al. Q-transformer: Scalable offline reinforcement learning via autoregressive q-functions. In Conference on Robot Learning , pages 3909-3928. PMLR, 2023. [4] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Dif- fusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137 , 2023. [5] Open X-Embodiment Collaboration et al. Open X- Embodiment: Robotic learning datasets and RT-X mod- els. https://arxiv.org/abs/2310.08864, 2023.","citing_arxiv_id":"2406.02523"},{"n":1,"role":"method","polarity":"use_method","paper_title":"3D Diffuser Actor: Policy Diffusion with 3D Scene Representations","primary_cat":"cs.RO","context_text":"Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V . Macua, S. Z. Tan, I. Momennejad, K. Hofmann, and S. Devlin. Imitating human behaviour with diffusion models, 2023. [6] C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 9 [7] M. Reuss, M. Li, X. Jia, and R. Lioutikov. Goal-conditioned imitation learning using score-based diffusion policies. arXiv preprint arXiv:2304.02532, 2023. [8] A. Mandlekar, F. Ramos, B. Boots, L. Fei-Fei, A. Garg, and D. Fox. IRIS: implicit reinforcement without interaction at scale for learning control from offline robot manipulation data. CoRR,","citing_arxiv_id":"2402.10885"},{"n":1,"role":"method","polarity":"use_method","paper_title":"Training Diffusion Models with Reinforcement Learning","primary_cat":"cs.LG","context_text":"Salimans, 2021) to produce perceptually high-quality results. CFG involves jointly training the diffusion model on conditional and unconditional objectives by randomly masking out the context c during training. The conditional and unconditional predictions are then mixed at sampling time using a guidance weight w: ˜ϵθ(xt, t, c) = wϵθ(xt, t, c) + (1 − w)ϵθ(xt, t) (3) where ϵθ is the ϵ-prediction parameterization of the diffusion model (Ho et al., 2020) and ˜ϵθ is the guided ϵ-prediction that is used to compute the next denoised sample. For reinforcement learning, it does not make sense to train on the unconditional objective since the reward may depend on the context. However, we found that when only training on the conditional","citing_arxiv_id":"2305.13301"}]},"error":null,"updated_at":"2026-07-09T00:53:07.184417+00:00"},"identity_refresh":{"job_type":"identity_refresh","status":"succeeded","result":{"items":[{"title":"Qwen3 Technical Report","outcome":"unchanged","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","resolver":"local_arxiv","confidence":0.98,"old_work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e"}],"counts":{"fixed":0,"merged":0,"unchanged":1,"quarantined":0,"needs_external_resolution":0},"errors":[],"attempted":1},"error":null,"updated_at":"2026-07-09T00:53:35.686392+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion","claims":[{"claim_text":"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 12 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 dynam","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"ceive, modulate, and record internal grasping force, external interaction wrench, and tactile interaction during demon- stration. In this sense, OmniUMI does not merely measure physical variables; it organizes their acquisition around a human-aligned interface for natural interaction regulation. At the learning and deployment level, we extend dif- fusion policy (DP) [6] to incorporate multimodal observa- tions, including visual, tactile, and force-related signals, en- abling policies to reason j","claim_type":"extension","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"9 53.9 42.0 32.0 2.89 TaKSIE [13] WACV, 2025 Plan 1 90.4 73.9 61.7 51.2 40.8 3.18 HULC++ [26] ICRA, 2023 Plan 1 93.0 79.0 64.0 52.0 40.0 3.30 RoboTron-Mani [40] ICCV, 2025 3D input 1 94.7 80.3 65.1 51.4 39.0 3.31 DaDu-Corki-SW [9] ISCA, 2025 - 1 92.3 80.0 67.4 56.6 45.8 3.42 RoboUniView (default) [23] arXiv, 2024 - 195.4 82.768.5 56.4 46.1 3.49 DTP [7] RA-L, 2025 Plan 1 92.4 81.970.2 60.3 50.9 3.55 NFE = 4 methods LaDi-WM [10] arXiv, 2025 WM≈4 92.7 83.1 72.1 61.2 54.1 3.63 LightDP [38] ICCV, 202","claim_type":"baseline","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"grounding [9, 290], world modeling [291, 292], and temporal prediction [293]. Since web data primarily consistsofvision-languagepairsandismainlyusedtoenhancefoundationmodelcapabilities, whichfallsoutside the scope of our survey, we focus primarily on human video in our discussion. Representative human video datasets include Ego4D [236], EPIC-KITCHENS-100 [277], and Something-Something V2 [275]. Although these datasets do not contain action labels directly usable for policy learning, they capture","claim_type":"dataset","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"network with data annotated at the sub-step level. Consequently, our method can complete com- plex tasks, like laundry folding, without requiring an external high-level policy, making the entire framework more end-to-end and demonstrating significant potential. Diffusion models. Diffusion models [48, 49, 50] have emerged as the dominant approach in visual generation. The Diffusion Policy [30] successfully applies the diffusion model to robot learning, demonstrating its ability to model multimoda","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"[2] primarily relied on autoregressive or behavior-cloning policies that predict actions step by step. More recently, dual-system VLA architectures [1, 9] have emerged, decoupling high-level multimodal reasoning [19, 22] from low-level action generation. In these systems, a perception-language back- bone [19] produces contextual representations, while a dedicated action expert is often implemented as a diffusion [23] or flow-matching head that predicts multiple future actions in parallel [24, 25","claim_type":"method","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"models in current usage; both are most usefully read here as design boundaries rather than additional roles. 10.1 Generative world simulators Generative world simulators form a boundary case. Systems inspired by large-scale video generation, interactive environment generation, and physical-AI simulation platforms are often called world models because they produce plausible future observations or playable environments [12, 49, 64], but photorealistic or interactive generation is not identical to ","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Diffusion Policy: Visuomotor Policy Learning via Action Diffusion because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (10 contexts).","role_counts":[{"n":10,"context_role":"background"},{"n":7,"context_role":"method"},{"n":1,"context_role":"baseline"},{"n":1,"context_role":"dataset"},{"n":1,"context_role":"extension"}]},"error":null,"updated_at":"2026-07-09T00:52:57.233001+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion","claims":[{"claim_text":"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 12 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 dynam","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"ceive, modulate, and record internal grasping force, external interaction wrench, and tactile interaction during demon- stration. In this sense, OmniUMI does not merely measure physical variables; it organizes their acquisition around a human-aligned interface for natural interaction regulation. At the learning and deployment level, we extend dif- fusion policy (DP) [6] to incorporate multimodal observa- tions, including visual, tactile, and force-related signals, en- abling policies to reason j","claim_type":"extension","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"9 53.9 42.0 32.0 2.89 TaKSIE [13] WACV, 2025 Plan 1 90.4 73.9 61.7 51.2 40.8 3.18 HULC++ [26] ICRA, 2023 Plan 1 93.0 79.0 64.0 52.0 40.0 3.30 RoboTron-Mani [40] ICCV, 2025 3D input 1 94.7 80.3 65.1 51.4 39.0 3.31 DaDu-Corki-SW [9] ISCA, 2025 - 1 92.3 80.0 67.4 56.6 45.8 3.42 RoboUniView (default) [23] arXiv, 2024 - 195.4 82.768.5 56.4 46.1 3.49 DTP [7] RA-L, 2025 Plan 1 92.4 81.970.2 60.3 50.9 3.55 NFE = 4 methods LaDi-WM [10] arXiv, 2025 WM≈4 92.7 83.1 72.1 61.2 54.1 3.63 LightDP [38] ICCV, 202","claim_type":"baseline","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"grounding [9, 290], world modeling [291, 292], and temporal prediction [293]. Since web data primarily consistsofvision-languagepairsandismainlyusedtoenhancefoundationmodelcapabilities, whichfallsoutside the scope of our survey, we focus primarily on human video in our discussion. Representative human video datasets include Ego4D [236], EPIC-KITCHENS-100 [277], and Something-Something V2 [275]. Although these datasets do not contain action labels directly usable for policy learning, they capture","claim_type":"dataset","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"network with data annotated at the sub-step level. Consequently, our method can complete com- plex tasks, like laundry folding, without requiring an external high-level policy, making the entire framework more end-to-end and demonstrating significant potential. Diffusion models. Diffusion models [48, 49, 50] have emerged as the dominant approach in visual generation. The Diffusion Policy [30] successfully applies the diffusion model to robot learning, demonstrating its ability to model multimoda","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"[2] primarily relied on autoregressive or behavior-cloning policies that predict actions step by step. More recently, dual-system VLA architectures [1, 9] have emerged, decoupling high-level multimodal reasoning [19, 22] from low-level action generation. In these systems, a perception-language back- bone [19] produces contextual representations, while a dedicated action expert is often implemented as a diffusion [23] or flow-matching head that predicts multiple future actions in parallel [24, 25","claim_type":"method","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"models in current usage; both are most usefully read here as design boundaries rather than additional roles. 10.1 Generative world simulators Generative world simulators form a boundary case. Systems inspired by large-scale video generation, interactive environment generation, and physical-AI simulation platforms are often called world models because they produce plausible future observations or playable environments [12, 49, 64], but photorealistic or interactive generation is not identical to ","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Diffusion Policy: Visuomotor Policy Learning via Action Diffusion because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (10 contexts).","role_counts":[{"n":10,"context_role":"background"},{"n":7,"context_role":"method"},{"n":1,"context_role":"baseline"},{"n":1,"context_role":"dataset"},{"n":1,"context_role":"extension"}]},"error":null,"updated_at":"2026-07-09T00:53:35.688991+00:00"}},"summary":{"title":"Diffusion Policy: Visuomotor Policy Learning via Action Diffusion","claims":[{"claim_text":"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 12 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 dynam","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"ceive, modulate, and record internal grasping force, external interaction wrench, and tactile interaction during demon- stration. In this sense, OmniUMI does not merely measure physical variables; it organizes their acquisition around a human-aligned interface for natural interaction regulation. At the learning and deployment level, we extend dif- fusion policy (DP) [6] to incorporate multimodal observa- tions, including visual, tactile, and force-related signals, en- abling policies to reason j","claim_type":"extension","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"9 53.9 42.0 32.0 2.89 TaKSIE [13] WACV, 2025 Plan 1 90.4 73.9 61.7 51.2 40.8 3.18 HULC++ [26] ICRA, 2023 Plan 1 93.0 79.0 64.0 52.0 40.0 3.30 RoboTron-Mani [40] ICCV, 2025 3D input 1 94.7 80.3 65.1 51.4 39.0 3.31 DaDu-Corki-SW [9] ISCA, 2025 - 1 92.3 80.0 67.4 56.6 45.8 3.42 RoboUniView (default) [23] arXiv, 2024 - 195.4 82.768.5 56.4 46.1 3.49 DTP [7] RA-L, 2025 Plan 1 92.4 81.970.2 60.3 50.9 3.55 NFE = 4 methods LaDi-WM [10] arXiv, 2025 WM≈4 92.7 83.1 72.1 61.2 54.1 3.63 LightDP [38] ICCV, 202","claim_type":"baseline","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"grounding [9, 290], world modeling [291, 292], and temporal prediction [293]. Since web data primarily consistsofvision-languagepairsandismainlyusedtoenhancefoundationmodelcapabilities, whichfallsoutside the scope of our survey, we focus primarily on human video in our discussion. Representative human video datasets include Ego4D [236], EPIC-KITCHENS-100 [277], and Something-Something V2 [275]. Although these datasets do not contain action labels directly usable for policy learning, they capture","claim_type":"dataset","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"network with data annotated at the sub-step level. Consequently, our method can complete com- plex tasks, like laundry folding, without requiring an external high-level policy, making the entire framework more end-to-end and demonstrating significant potential. Diffusion models. Diffusion models [48, 49, 50] have emerged as the dominant approach in visual generation. The Diffusion Policy [30] successfully applies the diffusion model to robot learning, demonstrating its ability to model multimoda","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"[2] primarily relied on autoregressive or behavior-cloning policies that predict actions step by step. More recently, dual-system VLA architectures [1, 9] have emerged, decoupling high-level multimodal reasoning [19, 22] from low-level action generation. In these systems, a perception-language back- bone [19] produces contextual representations, while a dedicated action expert is often implemented as a diffusion [23] or flow-matching head that predicts multiple future actions in parallel [24, 25","claim_type":"method","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"models in current usage; both are most usefully read here as design boundaries rather than additional roles. 10.1 Generative world simulators Generative world simulators form a boundary case. Systems inspired by large-scale video generation, interactive environment generation, and physical-AI simulation platforms are often called world models because they produce plausible future observations or playable environments [12, 49, 64], but photorealistic or interactive generation is not identical to ","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Diffusion Policy: Visuomotor Policy Learning via Action Diffusion because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (10 contexts).","role_counts":[{"n":10,"context_role":"background"},{"n":7,"context_role":"method"},{"n":1,"context_role":"baseline"},{"n":1,"context_role":"dataset"},{"n":1,"context_role":"extension"}]},"graph":{"co_cited":[{"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","shared_citers":46},{"title":"OpenVLA: An Open-Source Vision-Language-Action Model","work_id":"3e7e65c5-5aed-4fe9-8414-2092bcb31cc7","shared_citers":43},{"title":"Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware","work_id":"6fe159e0-fa73-481a-88d4-4719c15140be","shared_citers":40},{"title":"RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control","work_id":"ff438a8a-8003-4fae-9131-acd418b3597b","shared_citers":34},{"title":"RT-1: Robotics Transformer for Real-World Control at Scale","work_id":"e11bda85-8531-46bc-a07f-d0ade3643ab1","shared_citers":31},{"title":"$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization","work_id":"d1ad7304-d09a-49bc-809e-846439f6aff9","shared_citers":29},{"title":"Octo: An Open-Source Generalist Robot Policy","work_id":"f9ca0722-8855-48c3-a27a-0eefb7e19253","shared_citers":28},{"title":"Open X-Embodiment: Robotic Learning Datasets and RT-X Models","work_id":"62f0fb6c-e6ae-4dc4-95a4-d9dd64b240e8","shared_citers":26},{"title":"Flow Matching for Generative Modeling","work_id":"6edb71c4-5d64-40af-a394-9757ea051a36","shared_citers":22},{"title":"Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success","work_id":"04f46bb3-4346-47e8-bf09-c75d91f96e87","shared_citers":18},{"title":"GR00T N1: An Open Foundation Model for Generalist Humanoid Robots","work_id":"e2db69c7-ee8a-4cb7-a761-7b8de1dfcf97","shared_citers":18},{"title":"Denoising Diffusion Implicit Models","work_id":"8fa2128b-d18c-405c-ac92-0e669cf89ac0","shared_citers":17},{"title":"FAST: Efficient Action Tokenization for Vision-Language-Action Models","work_id":"83a8f966-6cfa-4f21-81f3-87440aae238f","shared_citers":17},{"title":"LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning","work_id":"662203ad-084f-42c4-8e60-977b3173755b","shared_citers":16},{"title":"RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation","work_id":"12319725-bc7d-4c32-a229-ad270a7460bc","shared_citers":14},{"title":"SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics","work_id":"0c5e9314-5fa7-4613-ad12-605a71d561d2","shared_citers":14},{"title":"Do As I Can, Not As I Say: Grounding Language in Robotic Affordances","work_id":"037320f1-b0a9-4cbe-a639-bfb25409ce71","shared_citers":13},{"title":"World Action Models are Zero-shot Policies","work_id":"9a85fc69-74df-450e-94cd-69d186e9e830","shared_citers":13},{"title":"3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations","work_id":"bded01e1-c070-4537-a75a-ace4c75d0c95","shared_citers":12},{"title":"Decoupled Weight Decay Regularization","work_id":"07ef7360-d385-4033-83f7-8384a6325204","shared_citers":12},{"title":"Denoising Diffusion Probabilistic Models","work_id":"dc023f4e-7c79-471c-b713-deeb559ba010","shared_citers":12},{"title":"DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset","work_id":"13253de2-3d89-415c-8c2f-3adb25d4c337","shared_citers":12},{"title":"Planning with Diffusion for Flexible Behavior Synthesis","work_id":"38b2c635-b754-412a-a8f5-dfcf3e405c95","shared_citers":12},{"title":"What Matters in Learning from Offline Human Demonstrations for Robot Manipulation","work_id":"6a4c95c5-540e-4854-946d-c7c8a6c540ba","shared_citers":12}],"time_series":[{"n":5,"year":2023},{"n":8,"year":2024},{"n":7,"year":2025},{"n":83,"year":2026}],"dependency_candidates":[{"n":1,"role":"method","polarity":"use_method","paper_title":"Dynamic Execution Commitment of Vision-Language-Action Models","primary_cat":"cs.CV","context_text":"[2] primarily relied on autoregressive or behavior-cloning policies that predict actions step by step. More recently, dual-system VLA architectures [1, 9] have emerged, decoupling high-level multimodal reasoning [19, 22] from low-level action generation. In these systems, a perception-language back- bone [19] produces contextual representations, while a dedicated action expert is often implemented as a diffusion [23] or flow-matching head that predicts multiple future actions in parallel [24, 25, 26]. In practice, these systems often execute a predefined number of predicted actions before re-planning, with the execution horizon typically selected per task or benchmark [27, 28, 29]. For instance, π-0.5 [1] adopts different execution horizons across LIBERO subtasks [ 8] to balance performance and","citing_arxiv_id":"2605.11567"},{"n":1,"role":"baseline","polarity":"baseline","paper_title":"CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies","primary_cat":"cs.CV","context_text":"9 53.9 42.0 32.0 2.89 TaKSIE [13] WACV, 2025 Plan 1 90.4 73.9 61.7 51.2 40.8 3.18 HULC++ [26] ICRA, 2023 Plan 1 93.0 79.0 64.0 52.0 40.0 3.30 RoboTron-Mani [40] ICCV, 2025 3D input 1 94.7 80.3 65.1 51.4 39.0 3.31 DaDu-Corki-SW [9] ISCA, 2025 - 1 92.3 80.0 67.4 56.6 45.8 3.42 RoboUniView (default) [23] arXiv, 2024 - 195.4 82.768.5 56.4 46.1 3.49 DTP [7] RA-L, 2025 Plan 1 92.4 81.970.2 60.3 50.9 3.55 NFE = 4 methods LaDi-WM [10] arXiv, 2025 WM≈4 92.7 83.1 72.1 61.2 54.1 3.63 LightDP [38] ICCV, 2025 Distill 493.7 84.5 74.1 64.4 55.6 3.72 NFE≥10 methods RoboTron-Mani (DiT) [40] arXiv, 2025 3D input>1096.9 83.068.1 56.5 46.8 3.51 𝜋 ∗ 0.5 [11] CoRL, 2025 3B 10 90.4 78.1 67.961.2 54.73.52 MDT (default) [31] RSS, 2024 - 10 93.","citing_arxiv_id":"2604.24622"},{"n":1,"role":"method","polarity":"use_method","paper_title":"Mask World Model: Predicting What Matters for Robust Robot Policy Learning","primary_cat":"cs.RO","context_text":"for reproducibility. Timestep conditioning via AdaIN-style modulation.In- stead of standard adaptive normalization, MWM applies timestep-dependent scale and shift modulation after normal- ization(Huang & Belongie, 2017). 4 For a hidden activation x, we first normalize using RM- SNorm, then modulate: ¯x= RMSNorm(x), Modulate(¯x;s) = ¯x⊙ 1 +α(s) \u0001 +β(s) (7) Here, α(s) and β(s) are functions of the timestep embed- ding s, learned through a combination of a trainable scale- shift table and timestep projection. The operator ⊙ repre- sents element-wise multiplication. This design enhances stability when operating on normal- ized V AE latents by effectively handling their inherent vari- ability, while simultaneously preserving accurate timestep-","citing_arxiv_id":"2604.19683"},{"n":1,"role":"extension","polarity":"extend","paper_title":"OmniUMI: Towards Physically Grounded Robot Learning via Human-Aligned Multimodal Interaction","primary_cat":"cs.RO","context_text":"ceive, modulate, and record internal grasping force, external interaction wrench, and tactile interaction during demon- stration. In this sense, OmniUMI does not merely measure physical variables; it organizes their acquisition around a human-aligned interface for natural interaction regulation. At the learning and deployment level, we extend dif- fusion policy (DP) [6] to incorporate multimodal observa- tions, including visual, tactile, and force-related signals, en- abling policies to reason jointly over geometry and inter- action. On the deployment side, we translate policy out- puts into virtual targets and execute them via impedance- based control, enabling implicit and unified regulation of motion and force without explicit switching between con-","citing_arxiv_id":"2604.10647"},{"n":1,"role":"method","polarity":"use_method","paper_title":"Referring-Aware Visuomotor Policy Learning for Closed-Loop Manipulation","primary_cat":"cs.RO","context_text":"on the referring action P a: we simply force its transla- tion component atrans to coincide with the referring point P, while leaving the rotation component arot free to be optimized by the policy model. We then implement our trajectory-steeringstrategy through a masked-denoising process (Tseng et al., 2023; Kim et al., 2023),i.e., zt =M ⊙A known + (1− M)⊙z t (6) where zt denotes the intermediate noisy action vector at diffusion timestep t, ⊙ represents the Hadamard (element- wise) product, Aknown is a known action vector used to steer the denoising, and M is a binary mask indicating the indices to replace within the full noisy trajectory. As stated in Sec. 3.2, this strategy is applied in two stages: (i)","citing_arxiv_id":"2604.05544"},{"n":1,"role":"dataset","polarity":"use_dataset","paper_title":"A Survey on Vision-Language-Action Models: An Action Tokenization Perspective","primary_cat":"cs.RO","context_text":"grounding [9, 290], world modeling [291, 292], and temporal prediction [293]. Since web data primarily consistsofvision-languagepairsandismainlyusedtoenhancefoundationmodelcapabilities, whichfallsoutside the scope of our survey, we focus primarily on human video in our discussion. Representative human video datasets include Ego4D [236], EPIC-KITCHENS-100 [277], and Something-Something V2 [275]. Although these datasets do not contain action labels directly usable for policy learning, they capture diverse human-object interactions, complex manipulation skills, and rich physical common sense, which are valuable sources of world knowledge. Their scale and diversity enable pretraining of temporal visual encoders and facilitate the learning of action token representations.","citing_arxiv_id":"2507.01925"},{"n":1,"role":"method","polarity":"use_method","paper_title":"DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control","primary_cat":"cs.RO","context_text":"network with data annotated at the sub-step level. Consequently, our method can complete com- plex tasks, like laundry folding, without requiring an external high-level policy, making the entire framework more end-to-end and demonstrating significant potential. Diffusion models. Diffusion models [48, 49, 50] have emerged as the dominant approach in visual generation. The Diffusion Policy [30] successfully applies the diffusion model to robot learning, demonstrating its ability to model multimodal action distributions. Subsequent research has further developed the Diffusion Policy [51, 52, 53, 7, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63] by applying it to 3D environments [64, 65, 66, 67], scaling its capabilities [29], improving its efficiency [68, 53],","citing_arxiv_id":"2502.05855"},{"n":1,"role":"method","polarity":"use_method","paper_title":"RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots","primary_cat":"cs.RO","context_text":"control at scale. In arXiv preprint arXiv:2212.06817 , 2022. [3] Yevgen Chebotar, Quan Vuong, Karol Hausman, Fei Xia, Yao Lu, Alex Irpan, Aviral Kumar, Tianhe Yu, Alexander Herzog, Karl Pertsch, et al. Q-transformer: Scalable offline reinforcement learning via autoregressive q-functions. In Conference on Robot Learning , pages 3909-3928. PMLR, 2023. [4] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Dif- fusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137 , 2023. [5] Open X-Embodiment Collaboration et al. Open X- Embodiment: Robotic learning datasets and RT-X mod- els. https://arxiv.org/abs/2310.08864, 2023.","citing_arxiv_id":"2406.02523"},{"n":1,"role":"method","polarity":"use_method","paper_title":"3D Diffuser Actor: Policy Diffusion with 3D Scene Representations","primary_cat":"cs.RO","context_text":"Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V . Macua, S. Z. Tan, I. Momennejad, K. Hofmann, and S. Devlin. Imitating human behaviour with diffusion models, 2023. [6] C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 9 [7] M. Reuss, M. Li, X. Jia, and R. Lioutikov. Goal-conditioned imitation learning using score-based diffusion policies. arXiv preprint arXiv:2304.02532, 2023. [8] A. Mandlekar, F. Ramos, B. Boots, L. Fei-Fei, A. Garg, and D. Fox. IRIS: implicit reinforcement without interaction at scale for learning control from offline robot manipulation data. CoRR,","citing_arxiv_id":"2402.10885"},{"n":1,"role":"method","polarity":"use_method","paper_title":"Training Diffusion Models with Reinforcement Learning","primary_cat":"cs.LG","context_text":"Salimans, 2021) to produce perceptually high-quality results. CFG involves jointly training the diffusion model on conditional and unconditional objectives by randomly masking out the context c during training. The conditional and unconditional predictions are then mixed at sampling time using a guidance weight w: ˜ϵθ(xt, t, c) = wϵθ(xt, t, c) + (1 − w)ϵθ(xt, t) (3) where ϵθ is the ϵ-prediction parameterization of the diffusion model (Ho et al., 2020) and ˜ϵθ is the guided ϵ-prediction that is used to compute the next denoised sample. For reinforcement learning, it does not make sense to train on the unconditional objective since the reward may depend on the context. However, we found that when only training on the conditional","citing_arxiv_id":"2305.13301"}]},"authors":[{"id":"e154a17f-60d9-4108-b473-30086aaeb366","orcid":null,"display_name":"Benjamin Burchfiel","source":"manual","import_confidence":0.72},{"id":"31a9bd4d-7ff4-4abd-ae50-4d238cdcf285","orcid":null,"display_name":"Cheng Chi","source":"manual","import_confidence":0.72},{"id":"f8532816-edc4-4aad-b739-881d427b2d7e","orcid":null,"display_name":"Eric Cousineau","source":"manual","import_confidence":0.72},{"id":"0d3d829c-ee2f-453a-bcd7-c5a168009f59","orcid":null,"display_name":"Siyuan Feng","source":"manual","import_confidence":0.72},{"id":"b222acc7-ffce-4cd8-9940-9375e32ace22","orcid":null,"display_name":"Yilun Du","source":"manual","import_confidence":0.72},{"id":"5ab9dce0-e5a8-4982-8f67-a1d384e9986c","orcid":null,"display_name":"Zhenjia Xu","source":"manual","import_confidence":0.72}]}}