Diffusion Policy

Follow this topic Ask about this topic

Diffusion policies learn to turn observations such as images, point clouds, and touch into action sequences, often representing several plausible ways to act rather than a single average behavior. Work commonly trains them through imitation of demonstrations, then adapts them with reinforcement learning, world models, or inference-time guidance; action chunking and multimodal sensing are widely used for manipulation, navigation, and whole-body control. Evaluations span simulated manipulation benchmarks and physical robots, including contact-rich and long-horizon tasks. A central challenge is retaining expressive, robust behavior while reducing denoising latency and improving feedback, safety, and generalization to new environments.

Latest papers

207 papers26 papers in the last four weeks, down 13% on the four weeks before. 0.2% of all new papers.

From 12 to 200 Demonstrations: How ACT, Diffusion Policy, SmolVLA and $π_{0.5}$ Scale on a Low-Cost Arm
Oct 27, 2026Simon CowlingImitation Learning Robot Imitation Learning
fysical.ai, Denmark
TAPESTRY: A Single Backbone for Video Generation and Robot Control
Oct 27, 2026Jeremy A. Collins, Seungjae Lee, Rachanon Wachakorn +4Vision-Based Robot Control Diffusion Policy
Georgia Institute of Technology. · Toyota Research Institute. · University of Maryland.
VaLoP: Value Gradient Lookahead and Pullback for Reward-Guided Diffusion
Oct 26, 2026Kwanyoung KimDiffusion Model Guidance Diffusion Guidance
HA-DiPO: Diffusion Policy Optimization for Heterogeneous Cooperative Transport
Oct 24, 2026Xuebin Li, Jinpeng Xiao, Zong Chen +4Multi-Robot Cargo Transportation Cooperative MARL
School of Mathematics, Harbin Institute of Technology, Harbin, China · State Key Laboratory of Intelligent Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China · Department of Electrical Engineering and Information Technology, University of Naples Federico II, Naples, Italy
Robotic Optical Coherence Tomography-Guided Geometric Modeling of Soft Tissue: 3D Representations for Data-Driven Closed-Loop Shape Control
Oct 23, 2026Chaoyu Zhang, Guangshen Ma, Mark DraelosDeformable Object Manipulation Robotic Manipulation
University of Michigan, Ann Arbor
Learning to Be Different: Component Specialization for Compositional Diffusion Policies
Oct 19, 2026Alec Laprevotte, Haonan Chen, Xiaoshen Han +1Diffusion Policy Multi-Task Learning
Harvard University
Compositional Humanoid Manipulation through Isomorphic Teleoperation and Independent Skill Routing
Oct 19, 2026Yiming Wang, Peixian Yang, Lifeng Zhuo +5Robotic Manipulation Humanoid Robotics
Shanghai Jiao Tong University. · The University of Hong Kong. · Xense Robotics. +1
SWAP: Test-Time Steerable Humanoid Whole-Body Collision Avoidance
Oct 19, 2026Zeji Yi, Guanqi He, Xiaofeng Guo +5Robot Collision Avoidance Humanoid Robot Control
Carnegie Mellon University · Amazon FAR
Robotic Spatiotemporal Editing: A Unified Interface for Controllable Action Generation via Masked Action Reconstruction
Oct 16, 2026Yanbang Li, Bu He, Yi Tian +6Robot Motion Generation Robotic Manipulation
Tsinghua University · CocoMatrix · Shanghai Qizhi Institute +1
Physical Time as a Missing Dimension of Robot Policy Memory Dynamics
Oct 14, 2026Yushen Liang, Yue Peng, Baosheng Jin +6Robot Policy Learning Memory-Augmented Agents
New York University Shanghai, Shanghai, China · University of Electronic Science and Technology of China, Chengdu, China
Steertuning Policies from a Few Human Videos
Oct 13, 2026Billy Yan, Isaac Tu, Kelly Lee +8Robot Policy Adaptation Data-Efficient Robot Learning
New York University
MOSAIC: Mode- and Support-Aware Inference-Time Control of Robot Diffusion Models
Oct 12, 2026Sabrina Bodmer, Rahel Rickenbach, René Zurbrügg +2Inference-Time Guidance Diffusion Guidance
Department of Mechanical and Process Engineering, ETH Zürich, Switzerland. · Automatic Control Laboratory, EPFL, Switzerland.
AUGUstus: Autonomous Driving VLA Model with Unified Generation and Understanding
Oct 12, 2026Yuqi Ye, Caigao Jiang, Zhuonan Wang +12NavSim Vision-Language-Action Models
Peking University · XPeng Motors · Fudan University
DR-APO: Decoupled-Representation Advantage-Guided Policy Optimization from Continuous Control to Embodied Navigation
Oct 12, 2026Runhua Zhang, Yingjie Niu, Chao LiPolicy Optimization RL for Generative Models
DEEP Robotics, Hangzhou, China · Zhejiang University, Hangzhou, China
DiffNash: Finding Nash Equilibrium in High-Dimensional Continuous Action Spaces with Diffusion Policies
Oct 11, 2026Haichuan Wang, Anagha Satish, Jiale Zheng +4Nash Equilibrium Computation Game-Playing Agents
Harvard University · Massachusetts Institute of Technology · Georgia Institute of Technology
Compositional Average Velocity Modeling for Efficient Diffusion Offline Reinforcement Learning
Oct 10, 2026Thanh Nguyen, Abu Hanif Muhammad Syarubany, Hwanhee Kim +1Offline RL Diffusion Policy
The authors are with the School of Electrical Engineering, KAIST, Daejeon, Republic of Korea.
CFG-Policy: Unified Diffusion Policy Learning Framework via Classifier-free Guidance
Oct 10, 2026Shutong Ding, Zhaoyang Liu, Yimiao Zhou +3Diffusion Policy Robotic RL
ShanghaiTech University, Shanghai, China · InstAdapt, Shanghai, China
Chronos Policy: Learning Robot Policies on Physical Time
Oct 9, 2026Changhao Wang, Jaden Clark, Shuran SongRobot Policy Learning Robot Learning
Stanford University
Control-Ready Uncertainty for Trajectory Diffusion
Oct 8, 2026Zhiwei Xue, Jia Yue Kam, Jinhang Qiu +5Uncertainty Calibration Probabilistic Trajectory Prediction
National University of Singapore. · LAAS-CNRS. · Smart Systems Institute, NUS.
SkillWeave: Weaving Heterogeneous Demonstrations into Long-Horizon Manipulation Skills
Oct 8, 2026Ryosei Tamura, Xiaoxiang Dong, Uksang Yoo +4Robotic Manipulation Long-Horizon Robotic Manipulation
Robotics Institute, Carnegie Mellon University, USA. · Department of Computer Science, Keio University, Japan. · College of Engineering, University of California, Berkeley, USA. +1
Immiscible Diffusion Policy: Preserving Multimodal Robot Actions through Label-Free Noise Assignment
Oct 7, 2026Xiao Zhang, Yuxin Chen, Zhixuan Liang +5Diffusion Policy Mode Collapse
University of California, Berkeley · Princeton University · The University of Texas at Austin
PEARS: Physical-Prior-Guided Efficient Adaptation via Failure Reasoning and Diffusion Steering for Tactile Manipulation
Oct 6, 2026Kun Song, Yiming Wang, Yilin Chen +5Robot Policy Adaptation Robotic RL
The University of Hong Kong · Shanghai Jiao Tong University · Xense Robotics +1
Reachability-Aware Diffusion Policy Optimization
Oct 5, 2026Hikmet Simsir, Kutay Demiray, Ozgur S. OguzConstrained RL Diffusion Policy
Department of Computer Engineering, Bilkent University
The Unexpired Plan: A Free Monitor for Accelerated Diffusion Policies
Oct 5, 2026Yi Zhao, Sebastian SchererRobotics Diffusion Policy
AirLab, The Robotics Institute Carnegie Mellon University, Pittsburgh, PA, USA
Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation
Oct 2, 2026Lik Hang Kenny Wong, Yiyao Ma, Xiu-Shen Wei +5Robot Imitation Learning Robot Learning
Department of Computer Science and Engineering, The Chinese University of Hong Kong · School of Computer Science and Engineering, Southeast University
Diffusion-2BC: Hybrid Diffusion and Regression Training for Offline Behavior Cloning in Autonomous Driving
Sep 29, 2026Bruno Maciel Machado, Eric Aislan AntoneloEnd-to-End Autonomous Driving Diffusion Policy
Department of Automation and Systems Engineering, Graduate Program in Automation and Systems Engineering, Federal University of Santa Catarina, Florianópolis, Santa Catarina, Brazil
DROM: A Language-Guided Diffusion Framework for Multi-Skill Robotic Manipulation
Sep 29, 2026Vincenzo Pomponi, Rocco Felici, Paolo Franceschi +5Long-Horizon Robotic Manipulation Robot Skill Learning
Institute of Systems and Technologies for Sustainable Production (ISTePS), Department of Innovative Technologies, University of Applied Science and Arts of Southern Switzerland (SUPSI), Via la Santa 1, Lugano, CH-6900, Ticino, Switzerland. · Istituto Dalle Molle di studi sull’intelligenza artificiale (IDSIA), Department of Innovative Technologies, University of Applied Science and Arts of Southern Switzerland (SUPSI), Via la Santa 1, Lugano, CH-6900, Ticino, Switzerland. · Mechanical Department, Politecnico di Milano (PoliMi), Via Giuseppe Candiani, Milan, 20155, Italy. +1
EquivDP3: A SIM(3)-Invariant Point-Cloud Encoder for Data-Efficient Humanoid Loco-Manipulation
Sep 29, 2026Abu Hanif Muhammad Syarubany, Chang D. YooHumanoid Loco-Manipulation Point Cloud Learning
School of Electrical Engineering Korea Advanced Institute of Science & Technology (KAIST) Daejeon, Republic of Korea
TLC-DiT: Task-Aligned Local Visual Conditioning for Robust Multitask Robot Manipulation
Sep 28, 2026Xianbo Cai, Hideyuki Ichiwara, Zihang Wang +2Language-Conditioned Robot Manipulation Multi-Task Robotic Manipulation
Department of Intermedia Art and Science, Waseda University, Tokyo, Japan · SB Intuitions Corp., Tokyo, Japan · Department of Electronic and Physical Systems, Waseda University, Tokyo, Japan +2
Principal Steering Subspaces for Online Adaptation of Frozen Generative Robot Policies
Sep 27, 2026Jialeng Ni, Nathan Zhao, Kunpeng SongRobot Policy Learning Reinforcement Learning
University of Michigan, Ann Arbor, MI, USA. · XPENG Robotics, Santa Clara, CA, USA.

Oct 27, 2026J/K move · Enter open · S save

From 12 to 200 Demonstrations: How ACT, Diffusion Policy, SmolVLA and $π_{0.5}$ Scale on a Low-Cost Arm

Simon Cowling

fysical.ai, Denmark

Imitation Learning Robot Imitation Learning Robotic Manipulation Robotic Policy Evaluation Diffusion Policy

Open paper AskPDF arXiv

We trained four widely used imitation-learning policies, ACT, Diffusion Policy, SmolVLA and

π_{0.5}

, on nested subsets of 12, 25, 50, 100 and 200 teleoperated demonstrations of one pick-and-place task with distractor objects, using one fixed training recipe per policy. Every trained model was then run on a physical SO-101 arm on a fixed set of object placements: 1,340 scored episodes in 126 blocks. ACT, trained from scratch, rose from 1 success in 40 episodes at 12 demonstrations to 37 in 40 at 100.

π_{0.5}

with its action expert fine-tuned was the best policy when data was scarce (16 of 40 at 12 demonstrations, 28 of 40 at 25) and then stayed near 70%; from 100 demonstrations ACT was ahead, by a margin this study does not fully resolve. SmolVLA and Diffusion Policy peaked near 50%, and both were affected by a failure to start: SmolVLA's arm never started moving in 47% of episodes, and Diffusion Policy hardly ever started until an idle lead-in was trimmed from the demonstrations. For

π_{0.5}

the fine-tuning recipe mattered more than the amount of data: low-rank adaptation succeeded in 4 of 110 episodes, against 130 of 200 for action-expert fine-tuning. From the joint and camera recordings we place every episode on a four-stage ladder (started, reached the ball, carried it, delivered it) and show that for ACT more data moves failures from reaching, to grasping, to none, while the failures

π_{0.5}

retains are in grasping. We also measure how weak a ten-episode block is as a measurement (the same checkpoint tested twice differs by 2 successes in 10 on average), separate seed-to-seed variation from that noise, and audit the human scoring against the recordings. Training length was fixed within each recipe, so the amount of data and the number of times it was seen change together; we state which conclusions survive that confound and which do not.