NeurIPS 2026 Workshop
Call for Papers
We invite submissions to RoboPAD: Post-Training Adaptation of Robot Foundation Models, a NeurIPS 2026 workshop in Paris, France. RoboPAD is a non-archival workshop designed to be a discussion venue rather than a formal proceedings track.
RoboPAD focuses on post-training as an algorithmic stage after pretraining: the closed-loop process of adapting, correcting, evaluating, and improving robot foundation models under real-world constraints. The central question is: after a robot foundation model has been pretrained, how should it be adapted, corrected, personalized, evaluated, and safely improved?
Accepted submissions are intended for presentation and community discussion. Accepted work does not appear in the official NeurIPS proceedings and does not preclude future archival submission.
One submission will receive the Best Paper Award, with an award certificate and recognition during the workshop. Additional Best Paper Runner-Up and Highlight Papers may also be recognized. All accepted papers are eligible; the final awardee is selected by the area chairs and organizers based on technical quality, novelty, clarity, and fit to the workshop theme.
Eligibility & What We Accept
RoboPAD broadly welcomes any work related to post-training adaptation of robot foundation models. Because the workshop is non-archival and discussion-oriented, we intentionally keep eligibility broad—authors are expected to respect the policies of other venues involved.
- Work in progress. Early-stage results, negative results, open problems, and discussion papers are all welcome.
- Papers under review elsewhere. Concurrent submission to conferences, journals, or other workshops is allowed.
- Already published or recently accepted papers. Papers previously published or accepted at archival venues (main conference, journal, prior workshop) may be resubmitted to RoboPAD for discussion and presentation, provided the authors confirm this is compatible with the original venue policy.
- Preprints, arXiv papers, and open-source / benchmark releases. Standalone benchmark, data, evaluation, or position contributions are all in scope.
- Contribution types. We welcome technical, empirical, theoretical, benchmark, and perspective / discussion submissions.
Author responsibility: Authors are responsible for ensuring that submission to RoboPAD is compatible with the policies of any other venue involved. If a paper is accepted for presentation elsewhere and the authors are unsure, they should disclose the situation at submission time and resolve conflicts with the other venue chairs directly.
Submission Format
RoboPAD uses a unified page upper bound aligned with the standard NeurIPS conference content-page limit. Because the workshop is non-archival, we intentionally do not force a single short length: authors may submit shorter workshop-style papers, or longer more complete technical papers, up to the same shared page cap.
Paper length: Up to 9 pages of main content, including all text, figures, and tables. References and appendix are not counted toward this page limit and are unlimited. Shorter workshop-style submissions (e.g., 4–7 pages) are fully welcome; the 9-page cap is an upper bound, not a target length.
Template: The official NeurIPS 2026 LaTeX style files, prepared in workshop format (\usepackage[dblblindworkshop]{neurips_2026} or the equivalent workshop style). Submissions that exceed the 9-page content-page limit will be desk-rejected without review.
References and appendix: References and appendix are unlimited. The appendix may be included in the same PDF as the main submission, directly after the references. Reviewers are not required to read the appendix; key technical claims and experiments should be placed in the main 9-page body.
Checklist: The NeurIPS main-conference paper checklist is not required for RoboPAD workshop submissions. A short Limitations / Broader Impact discussion is still encouraged but not mandatory.
Review mode: Double-blind. Submissions must be anonymized. Do not include author names, affiliations, or obvious self-citations without clearly anonymizing them in the review copy (e.g., “Anonymous et al.” or by omitting names and institutional links in figures / acknowledgments).
Code, data, and supplementary videos: Optional but strongly encouraged. You may link to anonymized repositories or URLs in the paper, and include additional results in the appendix. Separate supplementary ZIP uploads are not required for the workshop submission.
Submission & Review Process
Submissions are managed on OpenReview. We follow a lightweight review process designed for a workshop, while still maintaining double-blind review and conflict-of-interest enforcement.
OpenReview profiles: All authors and reviewers must have an OpenReview profile by the submission deadline. New profiles created without an institutional email may take time to activate; we recommend setting up profiles well in advance.
Review pipeline: Each submission receives at least two to three reviews; meta-reviews and final acceptance decisions are made by the RoboPAD area chairs and organizing committee.
LLM policy (authors): Authors may use LLM-based tools in accordance with the NeurIPS 2026 author policy. Authors remain fully responsible for the correctness and originality of all submitted content. Any substantive use of LLMs as part of the research methodology should be clearly disclosed.
LLM policy (reviewers): Reviewers and area chairs may not upload submissions or submission content to external LLM services, because this risks leaking confidential material.
Post-acceptance archival status: RoboPAD is non-archival. Accepted papers may be listed on OpenReview and the workshop website at the authors’ option. Accepted submissions are not part of the official NeurIPS proceedings. Authors remain free to update their arXiv versions and resubmit elsewhere in the future.
In-person presentation: At least one author of each accepted paper is expected to attend the workshop in person in Paris and present the work. Accepted submissions will be assigned either a poster slot or an oral slot (a subset of accepted papers). Registration and travel are the responsibility of the authors; acceptance does not include NeurIPS registration or travel funding.
Topics of Interest
The six themes below organize the main algorithmic problems in robot-foundation-model post-training. We welcome submissions on any related topic that engages substantively with adaptation, evaluation, and improvement after pretraining.
1. Feedback- and intervention-driven adaptation
- Learning from human corrections, preferences, teleoperation, and shared autonomy
- Online interventions, failure recovery, and interactive policy improvement
- Preference learning and human-in-the-loop specialization after pretraining
2. Policy optimization and specialization after pretraining
- RL fine-tuning of VLAs and pretrained robot policies
- Offline-to-online improvement, action decoders, and in-context policy adaptation
- Fast specialization with limited real-world data
3. World-model- and simulation-driven adaptation
- Adapting latent dynamics, predictive models, and task-level models
- Simulation-generated data for post-training under changing environments
- Closed-loop refinement of world models and policies before real-world deployment
4. Reasoning, memory, and agentic robot adaptation
- Coupling deliberative multimodal reasoning with reactive control
- Long-term robot memory, task decomposition, and when-to-think decisions
- Agentic pipelines for correction, replanning, and recovery
5. Cross-embodiment and whole-body adaptation
- Humanoids, dexterous hands, mobile manipulators, and embodiment transfer
- Specialist distillation, sim-to-real adaptation, and safety-aware post-training
- Whole-body and multi-robot specialization after pretraining
6. Evaluation, safety, and robustness after adaptation
- Reproducible protocols, robustness under distribution shift, and recovery metrics
- Constraint-respecting fine-tuning, reliability, and long-horizon evaluation
- Benchmarks and reporting practices for post-training outcomes
Presentation & Accepted Papers
Accepted papers will be presented in poster sessions on the day of the workshop. A subset of papers will additionally be selected for oral / spotlight presentations with Q&A. All accepted papers are eligible for the Best Paper Award.
No formal camera-ready paper is required. Authors are welcome to keep their submission up to date on arXiv and to share slides / posters after acceptance.
Important Dates
| Call for papers | Now open — submission via OpenReview |
| Submission deadline | 10 September 2026 (AoE) |
| Author notifications | 29 September 2026 (AoE) |
| Workshop date | One day within 12–13 December 2026, Paris, France |
Questions?
For questions about submissions, please contact the organizing committee at Li_Shijie@a-star.edu.sg.
This workshop is part of NeurIPS 2026. For general conference information, visit neurips.cc.