“Oops, I Erred”

Call for Contributions

You can contribute your work as one of two submission types. Submit through our OpenReview page by October 2, 2026 October 5, 2026.

All submissions should use the CoRL submission template. On the template, substitute
"Submitted to the 10th Conference on Robot Learning (CoRL 2026). Do not distribute." with
"Submitted to the "Oops, I Erred" - Workshop at CoRL 2026. Do not distribute."

(a) General Submission

We welcome paper submissions for algorithms, methods, systems, or analysis that include, support, and analyse the usage of imperfect and sub-optimal data for physical AI. These submissions will have a max page limit of 8 pages, excluding references and appendix. All submissions are non-archival. We invite contribution focusing on the topic of imperfect data in robot learning which includes but is not limited to the following:

• Collecting robotic datasets of failures, near-misses, out-of-distribution events, and other imperfect data at scale, including active learning and data collection in imperfect situations
• Classifying, annotating, formatting, and curating imperfect data for broad usability, including analysis that differentiates interesting failures from common ones
• Data sharing infrastructure, incentive structures, and standardization of large scale data that includes imperfect trajectories
• Literature reviews and analyses of what kinds of imperfect datasets exist, what is missing, and what is needed for the development of algorithms
• Algorithms that explicitly utilize failure and imperfect data to improve policy learning, world models, reward models, or other forms of modeling
• Algorithms for settings where imperfect data dominates success data
• Theoretical contributions and empirical investigations on understanding imperfect data, its utility, and the value (or lack thereof) of particular forms of imperfect data
• Analysis of failures with humans-in-the-loop, including trustworthy control algorithms for safe deployment when failure is critical

In general, we discourage submission of papers that have been accepted to an archival venue, especially those accepted to CoRL. We reserve the right to desk-reject papers that are already being presented at the main CoRL conference.

(b) Data Submission

In this workshop, we will trial-run a data contribution stream. In this stream, you can submit data for any physical AI domain along with a short paper (2-4 pages excluding references and appendix) describing how the data was collected, what the interesting characteristics in the data are, and how you think this data/data collecting technique can be useful to the community. Each submission should describe the data's origin, use-case, and scope, with documentation that allows users to analyse and use the data.

We recommend (but do not require) following the Oopsie formatting standards for contributors submitting manipulation data. Moreover, you are welcome to collect data in any physical AI domain, real or simulated and not limited to manipulation.

Data contributions are not automatically part of the Oopsie-Data project. If you would like, you can additionally register your data to be included in the Oopsie-Dataset, if the collected data meets the requirements (embodiment, modality, etc). Read more about contributing data to the dataset here. All contributors who have already submitted their data to Oopsie are welcome to submit a short paper describing how their data was collected.

Data submissions that are not part of the Oopsie-Dataset need to be hosted on a public repository. We highly recommend using HuggingFace. However, other options that are also often considered by for example NeurIPS datasets are Harvard Dataverse, Kaggle, or OpenML. Your institution might also have a partnership with Zenodo or Dryad that you can leverage for larger datasets.

Irrespective of hosting platform, clear metadata and documentation will be a reviewing criterion. Examples of reasonable documentation include using the Oopsie-data robot profile or more generally data cards to document your submission data.

Some questions you can address using your short papers:

• How to collect data interesting for physical AI?
• What are the various characteristics that should be expected in a dataset?
• What makes data useful?
• Any other question that can explain your data collection strategy.

Data Awards

To reward the often-invisible labor of dataset creation, we extend award mechanisms normally reserved for papers to datasets: participants and panelists vote for their favorite data contributions during a live session. The categories you can submit to are:

• Oopsie, I didn't mean that (most unexpected failure)
• Oopsie, almost made it (closest failure)
• Oopsie, E-stop required (most dangerous failure)
• Oopsie, don't laugh at me (funniest failure)

It does not matter if your submission is in the data or general track. However, to be considered for an award, you must make the data that was used in the submission publicly available. It needs to be part of the workshop paper submission.

Important Dates (All times are in AOE)

Review and Selection Process

• All reviews will be double-blind
• All accepted papers will have posters at the workshop
• Papers will have non-archival status
• At least one co-author of an accepted paper is expected to attend the workshop
• All accepted papers will be made public on OpenReview