RoboGen: 3D World Generation for Robot Learning and Autonomous Systems

Overview

Workshop Details

Welcome to the IROS 2025 RoboGen Workshop on World Understanding and Generation!

RoboGen focuses on 2 aspects: multimodal world understanding and generation, because we believe these two problems are tightly bound, only if understanding is done right, then the generation is reasonable. Recent advances in 3D Vision (NeRF, Gaussian Splatting) and multimodal foundation models (LLM, VLM, diffusion and flow-based model) are transforming robotics by enabling the creation of high quality data for training, testing, and validation. Despite impressive capabilities of learning-based robotics in embodied AI, autonomous driving, unmanned aerial navigation, progress in generalizable systems remains constrained by the fundamental challenge of data acquisition.

Building towards AGI, RoboGen brings together researchers and industry experts to address this data bottleneck for scaling law. The goal is to develop innovative approaches that leverage recent advances in 3D scene generation and understanding methods to . Our workshop emphasizes practical applications and solutions to long-tail data problems in challenging robotics scenarios, with the goal of empowering 3D deep learning systems for real-world deployment.

The RoboGen workshop aims to advance 3D world generation for robotics with four key objectives:

(1) Understanding: advance multimodal world understanding and spatial-temporal reasoning through vision-language models,
(2) Representation: explore video diffusion and flow-based models conditioned on sensor geometry cues for physical 3D world representation,
(3) Action: democratize access to high-quality synthetic data generation to train robust vision-language-action models.
(4) Scaling: enable practical deployment of generalizable robot learning systems overcoming the data scaling law bottleneck.

Workshop Features

The workshop will feature the following events and activities to encourage discussion and participation. We are looking for industry partners to showcase their products and services. If you are interested in sponsoring the workshop, please contact us at robogen-iros@googlegroups.com.

Industry-Academia Keynote Sessions
Leading researchers and practitioners from both academia and industry will present cutting-edge advancements in multimodal foundation models and 3D world generation methods for robotics applications.
Oral Presentations
Selected paper presentations highlighting impactful approaches to data generation, artifact mitigation, and domain-specific applications of neural rendering and generative models for robotics.
Interactive Poster & Demo Session
A dedicated session where student researchers can demonstrate their methods and techniques, fostering discussions and collaborations. This will provide an opportunity for junior researchers to directly interact with senior researchers.
Expert Panel Discussion
A focused discussion on bridging the gap between theoretical advancements in neural rendering and practical robotics applications.
Award Ceremony
Awards will be presented in multiple categories including Best Paper, Best Poster, Most Impactful Open-Source Project, and Most Engaging Speaker to encourage participation and recognize outstanding contributions.

Mentor 1:1 Session
Beyond formal talks and breaks, we're facilitating personalized 15-30 minute one-on-one sessions between junior researchers and keynote speakers or organizing committee members during IROS. Prior to the conference, RoboGen workshop participants will receive a signup form to indicate their interests. This creates valuable mentorship opportunities, connecting junior researchers with senior experts for both technical guidance and career development.

Paper Submission

Call for Papers

Important Note

All submissions are non-archival, which allows authors to submit to other conferences and journals in the future. In addition, we welcome papers that have been submitted to or accepted by other venues, as well as highly impactful open-source projects .

📚 Topics

Topics of this workshop include but are not limited to:

Vision-language conditioned image/video generation
Embodied robot training through exploration in interactive synthetic environments
Multimodal world understanding with spatial-temporal reasoning
Long-horizon AV/robotics video reasoning using chain-of-thought and reinforcement learning
Post-training multimodal LLMs to reason about the physical world
Inference-time scaling of general-purpose world models for cost-effective robot deployment
Simultaneous sensor and traffic simulation for novel environment adaptation in robotics
Multi-modal sensing integration (radar, lidar) with diffusion-based generative world models
4D dynamic scene composition with NeRF and 3D Gaussian Splatting using cost-effective multi-modal sensors
Robust perception in adverse environments (low-light, long-range, sparse view) using neural rendering
Robot simultaneous localization and mapping (SLAM) with neural or Gaussian scene representation

📅 Important Dates

Event	Date
Submission open	June 1st, 2025, 11:59 PM AOE
Submission due	August 31st, 2025, 11:59 PM AOE
Notification to authors	September 7th, 2025, 11:59 PM AOE
Camera Ready	October 1st, 2025, 11:59 PM AOE
Workshop Date	Oct 24th, 2025

📝 Submission Format

Authors may submit either (1) extended abstracts (2–4 pages) , which is suitable for preliminary work-in-progress or already-published results; (2) full papers (6–8 pages) that present original, detailed research contributions (methodology, experiments, and analysis).

Formatting Requirements:
All submissions must follow the IROS/IEEE double column format with page limits that include references, appendices, and any additional material. Papers should be submitted as a single PDF file (up to 6MB). Official formatting templates are available here.

Submission Portal:
RoboGen @ OpenReview

To recognize and encourage outstanding contributions, RoboGen will present awards in the following categories. All awards will be selected by the organizing committee:

• Best Paper: Three best paper candidates will be invited for oral spotlight presentations, with one selected as the winner
• Best Poster: Selected from all accepted papers presented in the poster session
• Best Open-Source: Open to novel datasets, projects or already published open-source works of great impact in the field
• Best Speaker: Awarded to the most engaging speaker across all presentations

Additional Information:
    • All submissions will be peer-reviewed
    • Optional supplementary materials (videos, images, code) may be submitted as a separate ZIP file
    • Accepted papers and abstracts will be made publicly available on this website
    • Authors will present their work in either oral or poster sessions

Speakers

Invited Speakers

Program

Workshop Schedule (Tentative)

All invited talks, oral presentations and the panel discussion will take place in-person at IROS in Hangzhou, China with support for remote participation.