Open calls for contributions
We invite laboratories, individual researchers, and teams to contribute manipulation datasets using HiveBoard or develop attachments for two benchmark extensions.
| Open call | Scope |
|---|---|
| Learning datasets | Demonstrations and robot rollouts with states, actions, camera observations, calibration, and task outcomes for learning-based methods |
| Activities of daily living (ADL) | Mechanisms representing everyday manipulation tasks, such as operating closures, dispensing items, or handling household objects |
| Bimanual manipulation | Tasks that require coordinated action by two hands or end-effectors |
Contribute a learning dataset
Record robot states, executed commands, and camera streams while performing HiveBoard tasks. Include timestamps, calibration, an episode index, and a loading example. There is no fixed episode count or requirement to cover all conditions; larger collections are encouraged. Contact the organizers before collection to agree on tasks, signals, and formats. See the dataset call for requirements and direct email submission.
We recommend DataHive for collection, annotation, validation, and upload. Labs performing a benchmark evaluation are encouraged to record learning data during those trials; collecting a dataset remains optional.
Substantial accepted datasets may establish eligibility for authorship on future dataset papers that use the contribution. Smaller contributions may be acknowledged. See the dataset credit policy.
Propose a new attachment
Contact Ricardo V. Godoy before starting detailed design or fabrication. Send a short task description, a sketch or reference image, and the intended call. The authors will discuss whether the idea fits the benchmark, overlaps with an existing attachment, and can be tested consistently across platforms.
An initial discussion establishes the scope; acceptance follows review of the working attachment and its files.
What to contribute
Submit a functional physical prototype, editable CAD, printable files, an articulated simulation model, assembly instructions, task and reset definitions, and validation evidence. The submission checklist specifies the complete package.
Contributing at least one accepted, functional attachment with the complete package can establish eligibility for authorship on future papers that use the contribution. Smaller contributions may be acknowledged. Read the credit policy before starting.
The ADL and bimanual calls concern new benchmark extensions. The learning-dataset call has no fixed episode count. Formal benchmark evaluations still require 13 conditions with five trials each, for 65 trials total. Proposed attachments do not enter the current evaluation until a revised protocol is released.