IROS 2026 WORKSHOP · PITTSBURGH

Physical World Models for Scaling Embodied AI (PWMS2026)

From Scalable Physical Experience to Predictive Embodied Action.

Accepted Papers WorldArena 2.0 Challenge View Program
DateThursday, October 1, 2026
Time1:30 PM – 5:30 PM (Half Day)
LocationDavid L. Lawrence Convention Center, Pittsburgh, PA, USA

About the Workshop

Embodied AI is entering a scaling moment, but scaling embodied systems requires more than larger policies or more robot demonstrations. It requires diverse, interaction-rich, and physically grounded experience that can be generated, structured, evaluated, and reused across tasks, embodiments, and environments. This workshop explores Physical World Models as a new scaling infrastructure for embodied AI, turning human and robot experience into scalable simulation, synthetic data, predictive rollouts, and action-relevant representations.

We organize the workshop around two tightly connected directions. Physical World Models as Data Engines asks how human videos, robot demonstrations, cross-embodiment data, Real2Sim2Real pipelines, 3D/4D world generation, visuo-tactile sensing, and data-quality benchmarks can create reliable experience for robot learning. World Action Models for Embodied Decision-Making asks how video prediction, geometry grounding, latent dynamics, touch, long-horizon reasoning, policy evaluation, and self-improving agents can bridge predicted futures and executable actions in real-world embodied systems.

Topics of Interest

Topics include, but are not limited to:

Direction I — Physical World Models as Data Engines

From human and robot experience to scalable simulation, synthetic data, and evaluation.
  • Learning from human videos, demonstrations, and human-object interaction: extracting task goals, object affordances, action segments, contact events, and interaction priors from in-the-wild and egocentric human demonstrations for physical world model training and robot policy learning.
  • Cross-embodiment robot data scaling and transfer: building multi-robot data mixtures, shared state/action representations, embodiment-aware metadata, and action retargeting methods to transfer experience across arms, grippers, mobile manipulators, humanoids, and other robot platforms.
  • Real2Sim2Real and digital twins: reconstructing editable simulation environments from real-world observations, generating controllable rollouts, and using real-world feedback to refine simulation and world models.
  • 3D/4D world modeling and generation: modeling scene geometry, object states, dynamic changes, affordances, occlusions, contact regions, and uncertainty for physically grounded data generation.
  • Synthetic data generation for policy learning and evaluation: generating controllable demonstrations, simulated rollouts, scene and object-state variations, initial states, rare events, failure cases, and evaluation episodes to train, stress-test, and compare robot policies.
  • Visuo-tactile data generation for contact-rich manipulation: generating and annotating synchronized vision, tactile, force/torque, proprioceptive, and action signals for learning physical interactions.
  • Benchmarks and evaluation protocols for world-model-generated data: measuring how synthetic data, simulated rollouts, and real-synthetic data mixtures affect downstream robot policy performance across tasks, embodiments, and sensing modalities.

Direction II — World Action Models for Embodied Decision-Making

From world prediction to action generation, policy evaluation, and autonomous learning.
  • Video-based World Action Models: using video world models and video-action prediction to imagine future observations, task progress, and executable robot actions for manipulation, navigation, and whole-body control.
  • Geometry-aware World Action Models: grounding action generation in 3D/4D scene structure, object states, spatial relations, point flows, contact geometry, and multi-view consistency for physically reliable embodied action.
  • Latent World Action Models for efficient planning: learning compact latent dynamics, latent actions, visual subgoals, and action-conditioned representations for low-latency prediction, planning, and robot control.
  • Vision-tactile World Action Models for contact-rich manipulation: coupling visual, tactile, force/torque, proprioceptive, and action representations to predict contact transitions, slip, deformation, force evolution, and corrective actions for fine manipulation.
  • Long-horizon World Action Models for task-level reasoning: integrating future prediction with subgoal discovery, temporal abstraction, value estimation, risk awareness, and memory for multi-step manipulation, mobile manipulation, and embodied planning.
  • Predictive policy evaluation and safety-aware action selection: using world-model rollouts to estimate success, risks, constraint violations, and failure modes before selecting reliable actions for real-world execution.
  • Agentic and self-improving World Action Models: enabling embodied agents to learn from imagined rollouts, deployment feedback, interventions, failures, and recovery trajectories for autonomous data collection, policy improvement, and continual adaptation.

Invited Speakers

Jiajun Wu

Jiajun Wu

Stanford University, USA
Abhinav Valada

Abhinav Valada

University of Freiburg, Germany
Rudra P.K. Poudel

Rudra P.K. Poudel

Toshiba Europe, UK
Ding Zhao

Ding Zhao

Carnegie Mellon University, USA
Katerina Fragkiadaki

Katerina Fragkiadaki

Carnegie Mellon University, USA

Program

TimeSession
13:30 – 13:35Opening Remarks — Haibao Yu (CMU, USA)
13:35 – 14:00Keynote 1 — Jiajun Wu (Stanford University, USA)Building Visual and Physical Intelligence Through Code
14:00 – 14:25 Keynote 2 — Rudra P.K. Poudel (Toshiba Europe, UK) Local World Models: Adapting Foundation Models for X
14:25 – 14:40Keynote 3 — Lian Ruan (X Square Robot)Building Robot Foundation Models for the Physical World
14:40 – 14:55Keynote 4 — Mingle Li (SimpleAI)From Experience to Intelligence: Closed-Loop Robot Learning for Deployable Robotics
14:55 – 15:15
Oral Presentations I (2 × 8 min)
  • Yaru Niu (Carnegie Mellon University)
  • Haoyi Niu (University of California, Berkeley)
15:15 – 15:30Coffee Break and Poster Session
15:30 – 15:55Keynote 5 — Abhinav Valada (University of Freiburg, Germany)World Models for Open-World Autonomy
15:55 – 16:20Keynote 6 — Ding Zhao (Carnegie Mellon University, USA)Physical AI Safety: From Rare-Event Generation to Predictive Safety Evaluation
16:20 – 16:45Keynote 7 — Katerina Fragkiadaki (Carnegie Mellon University, USA)Scaling Robot Data Human-Free using Generative Planning and Foundation Models
16:45 – 17:05
Oral Presentations II (2 × 8 min)
  • Ray Muxin Liu (Carnegie Mellon University)
  • Binghong Chen (Carnegie Mellon University)
17:05 – 17:20Challenge Session — WorldArena 2.0 Challenge track results and winning-team presentations
17:20 – 17:30Best Paper and Best Poster Awards & Closing Remarks

Call for Papers

We invite original, unpublished work on Physical World Models and World Action Models for scaling Physical AI.

Non-archival workshop

Accepted papers will not appear in the IROS proceedings, and submission does not preclude future publication at other venues.

Submission Guidelines

01Research Fit

Scope

Physical World Models as scalable data engines and World Action Models for embodied decision-making.

02Submission

Format & Review

4–8 pages, excluding references, using the IROS workshop template; double-blind review.

03Contributions

Contribution Types

Technical and position papers, datasets, benchmarks, challenge reports, and negative results; reproducible artifacts are encouraged.

04Workshop

Presentation

All accepted papers must be presented as posters, with selected papers invited for oral spotlight presentations.

Paper & Poster Timeline

Submission Deadline
Review Deadline
Notification of Acceptance
Camera-ready
Poster Session at IROS 2026
Poster Session Award

Best Poster Award

Presented during the Poster Session at IROS 2026.

$500Prize

Submissions will be reviewed for relevance, originality, technical quality, clarity, and potential impact.

Dates are tentative and may be adjusted. Submissions will be handled via OpenReview.

Accepted Papers

20 papers

Non-archival workshop · Papers will not appear in the IROS proceedings.

[Oral] Learning Versatile Humanoid Manipulation with Touch Dreaming

Yaru Niu1,3, Zhenlong Fang1, Binghong Chen1, Shuai Zhou1, Revanth Krishna Senthilkumaran1,3, Hao Zhang1,2, Bingqing Chen3, Chen Qiu3, H. Eric Tseng2, Jonathan Francis1,3, Ding Zhao1

Affiliations
  • 1Carnegie Mellon University, Pittsburgh, USA
  • 2The University of Texas at Arlington, Arlington, USA
  • 3Bosch Center for Artificial Intelligence, Pittsburgh, USA
Project

[Oral] Egocentric Cross-Embodiment Manipulation with Embodiment Dreaming

Binghong Chen1,*, Yaru Niu1,*, Changwei Yao1,*, Shangtao Li1,*, Zhenlong Fang1, Revanth Krishna Senthilkumaran1,3, Shuai Zhou1, Yuemin Mao1, Yudong Luo1, Peiqi Yu1, Hao Zhang1,2, Bingqing Chen3, Chen Qiu3, H. Eric Tseng2, Changliu Liu1, Jonathan Francis1,3, Ding Zhao1

Affiliations
  • 1Carnegie Mellon University, Pittsburgh, USA
  • 2The University of Texas at Arlington, Arlington, USA
  • 3Bosch Center for Artificial Intelligence, Pittsburgh, USA

* Equal contribution.

Project

[Oral] A Recipe for Efficient Sim-to-Real Transfer in Manipulation with Online Imitation–Pretrained World Models

Yilin Wang1, Shangzhe Li2, Haoyi Niu3, Jiawei Fu1, Zhiao Huang4, Weitong Zhang2, Hao Su4

Affiliations
  • 1UC San Diego, USA
  • 2University of North Carolina at Chapel Hill, USA
  • 3UC Berkeley, USA
  • 4Sudo AI GmbH, Switzerland
Paper

[Oral] IFG: Internet-Scale Guidance for Functional Grasping Generation

Ray Muxin Liu1,*, Mingxuan Li1,*, Kenneth Shaw1, Deepak Pathak1

Affiliations
  • 1Carnegie Mellon University, Pittsburgh, PA, USA

* Equal contribution.

Project

[Oral] KnockGS: Interaction-Grounded Calibration of Physical Gaussian Representations

Chenchen Ge1,2,*, Hanwen Shen3,*, Bowen Jing1, Jiyuan Cai7, Xiaofeng Wang4,8, Hongsen Lei9, Weitao Zhou4,5, Dandan Zhang6, Haibao Yu1,10,†

Affiliations
  • 1Tuojing Intelligence
  • 2Southeast University
  • 3Stevens Institute of Technology
  • 4Tsinghua University
  • 5Simple AI
  • 6Imperial College London
  • 7Shanghai Jiao Tong University
  • 8GigaAI
  • 9Sun Yat-sen University
  • 10The University of Hong Kong

* Equal contribution. † Corresponding author.

Project

[Oral] Wild-World: Action-Conditioned RGB-D-Semantic World Modeling for Field Robotics

Zihang Wang, Yuhang Chu, Wenkai Zhu, Huateng Wu, shuyuan Li, Dong Kong

[Oral] Action Response Is Not Factual Direction: A Prospective Audit of Robot Video World Models

Yinuo Chen1

Affiliations
  • 1The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China
Paper

[Poster] Closed-Loop Planning in Latent World Models for Contact-Rich Manipulation: A Diagnostic Study on Push-T

Jiong Ding1, Xu Mohan1, Zhuoran Xie1, Qingyun Ye1, Chen Rui1, Fan Zhou1

Affiliations
  • 1Beihang University, Beijing, China
Project

[Poster] When Ensemble Disagreement Is Not a Physics Check: Stress-Testing World Models Under Contact-Parameter Shift

Krishna Harish

[Poster] Vision Language Grasping System Adapted for Digital Twin

Yanran Yi, Xiao Li

[Poster] Gaussian4XWAM: Gaussian Field Enhancement for World-Action Models

Zijian Zhang1, Yuqing Jiang1, Haibao Yu2

Affiliations
  • 1University of Chinese Academy of Sciences, Beijing, China
  • 2The University of Hong Kong, Hong Kong SAR, China
Project

[Poster] Towards Trustworthy Surgical Gesture Recognition with Uncertainty-Calibrated Kinematic World Models

Mu Xing1, Lan Wei1, Piotr Blaszyk1, Dandan Zhang1

Affiliations
  • 1Imperial College London, London, UK

[Poster] Bridging Semantics and Physics with Constrained LLMs for Safe and Trustworthy Robotic Manipulation*

Lan Wei1, Wenhao Hong1, Dandan Zhang1

Affiliations
  • 1Imperial College London, London, UK
Paper

[Poster] Prediction Fidelity Predicts Some Uses of a World Model and Not Others

Varun Puttagunta

[Poster] Anticipation-VLA: Solving Long-Horizon Embodied Tasks via Anticipation-based Subgoal Generation

Zhilong Zhang1,2, Wenyu Luo2, Haonan Wang2, Yifei Sheng2, Yidi Wang1,2, Hanyuan Guo2, Haoxiang Ren2, Xinghao Du1,2, Yuhan Che3, Tongtong Cao3, Lei Yuan1,2, Yang Yu1,2,†

Affiliations
  • 1National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China
  • 2School of Artificial Intelligence, Nanjing University, Nanjing, China
  • 3Department of Foundation Model, 2012 Labs, Huawei, Beijing, China
Paper

[Poster] Change-Localized Streaming Gaussian Spatial Memory for Long-Video 4D Reconstruction

Hanwen Shen2, Lei Yang3, Haibao Yu1,4

Affiliations
  • 1Tuojing Intelligence
  • 2Stevens Institute of Technology, Hoboken, USA
  • 3Nanyang Technological University, Singapore
  • 4The University of Hong Kong, Hong Kong
Code

[Poster] AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation

Mengfei Zhao9, Dihong Huang1,2, Yikai Tang2, Peihao Li2, Mingxuan Yan3, Ruiqi Zhuang1, Yanjia Huang4, Jie Wang1,5, Hai Zhai1, Tony Zhou1,6, Rui Zhang1,7, Zhexi Luo1,8, Yuchen Huang1,7, Jianfei Yang9, Jiachen Li3

Affiliations
  • 1Axis Robotics
  • 2University of California, Berkeley
  • 3Georgia Institute of Technology
  • 4Texas A&M University
  • 5Johns Hopkins University
  • 6University of Pennsylvania
  • 7University of Michigan
  • 8National University of Singapore
  • 9Nanyang Technological University
Paper

[Poster] Composing Unseen Behavior Combinations Across Worlds with Transplantable Dynamics

Abhinav Sadras1

Affiliations
  • 1Vellore Institute of Technology, Vellore, India
Project

[Poster] Decision-Level Auditing of Latent Robot World Models

Anikait Rana1, Siddhant Rana1, Aditya Rao Udupi1

Affiliations
  • 1Independent Researcher
Paper

WorldArena 2.0 Challenge

Benchmarking embodied world models from perceptual quality to interactive learning and real-world manipulation. Winning teams will present their solutions during the Challenge Session.

Visit Official Challenge Website
01Perception

Video Quality Evaluation

Visual and motion quality, content consistency, physics adherence, 3D accuracy, and controllability.

02Interactive Learning

World Model as RL Environment

Whether world models can serve as interactive environments for reinforcement learning and policy optimization.

03Physical Action

Real-World WAM Manipulation

Real-world manipulation performance in tactile WAM and vision-only WAM settings.

Challenge Timeline

Competition Opens
Final Submission Deadline
Final Results Released
Challenge Session at IROS 2026

Awards

Awards are presented separately for each competition track.

Track 01

Video Quality Evaluation

  1. 11st Prize
    Certificate + IROS Workshop Presentation
    $700
  2. 22nd Prize
    Certificate
    $400
  3. 33rd Prize
    Certificate
    $300
Track 02

World Model as RL Environment

  1. 11st Prize
    Certificate + IROS Workshop Presentation
    $1,400
  2. 22nd Prize
    Certificate
    $800
  3. 33rd Prize
    Certificate
    $600
Track 03

Real-World WAM Manipulation

  1. 11st Prize
    Certificate + IROS Workshop Presentation
    $1,400
  2. 22nd Prize
    Certificate
    $800
  3. 33rd Prize
    Certificate
    $600

Organizers

Our team spans the USA, UK, Singapore, China, and UAE, combining expertise in vision, robotics, and safety validation.

Haibao Yu

Haibao Yu

Carnegie Mellon University, USA
Spatial Intelligence, Generative Simulation, Embodied AI
Dandan Zhang

Dandan Zhang

Imperial College London, UK
Embodied Intelligence, Human-Robot Interaction, Medical Robotics
Lei Yang

Lei Yang

Nanyang Technological University, Singapore
Spatial Intelligence, World Models
Weitao Zhou

Weitao Zhou

Tsinghua University, China
Autonomous Driving, RL, Safety-Critical Planning
Jianing Qiu

Jianing Qiu

Mohamed bin Zayed University of AI, UAE
Foundation Models, Agents, Human-AI Interaction
Jiankai Sun

Jiankai Sun

Stanford University, USA
Embodied AI, Multimodal LLMs, World Models
Tianxing Chen

Tianxing Chen

The University of Hong Kong, China
Embodied AI, Robot Foundation Models, Manipulation
Lan Wei

Lan Wei

Imperial College London, UK
Diffusion-based Image Generation, VLA Models
Wen Fan

Wen Fan

Imperial College London, UK
Vision-based Tactile Sensing, Multimodal Robotic Manipulation
Walter Zimmer

Walter Zimmer

University of California, Los Angeles (UCLA), USA
Autonomous Driving, Intelligent Transportation Systems, Foundation Models
Chen Gao

Chen Gao

Tsinghua University, China
World model, Embodied intelligence, LLM/VLM agents
Yong Li

Yong Li

Tsinghua University, China
Artificial intelligence, Data science
Ping Luo

Ping Luo

The University of Hong Kong, China
Vision-Language Models, Autonomous Driving

Program Committee

Program Committee members

Yingjuan Tang

Yingjuan Tang

Zhengzhou University of Light Industry, China
Lingjun Zhang

Lingjun Zhang

Tsinghua University, China
Yuechen Luo

Yuechen Luo

Tsinghua University, China
Hanwen Shen

Hanwen Shen

Stevens Institute of Technology, USA
Ziying Song

Ziying Song

Nanyang technological University, Singapore
Yang Lyu

Yang Lyu

University of Michigan, USA
Yuner Zhang

Yuner Zhang

Carnegie Mellon University, USA
Zuo Yushen

Yushen Zuo

SimpleAI, China
Jinhao Zhang

Jinhao Zhang

Harbin Institute of Technology (Shenzhen), China
Junfan Zhu

Junfan Zhu

Saturday Robotics

Get in Touch

For submission questions, sponsorship inquiries, or program updates, contact the lead organizers directly.

yuhaibao94@gmail.com lei.yang@ntu.edu.sg

Venue

David L. Lawrence Convention Center
1000 Fort Duquesne Boulevard, Pittsburgh, PA 15222, United States

Held as part of IROS 2026 (September 27 – October 1, 2026). Room assignment will be announced in the official IROS program. The workshop is an in-person event.