Skip to content

Data and models

Public demonstration dataset

The original six-task demonstrations are available on Hugging Face: dancher00/WasserMan. The dataset retains all 504 collection attempts and the 480 selected demonstrations (76 training and 4 validation episodes per task), including lossless RGB, commands, measured state, physical traces and original audit records. The download is 35.69 GiB; one-task and one-episode downloads are supported. Follow the dataset card for checksummed download and restoration. This publication contains demonstrations; fitted models and complete policy evaluation packages have separate availability.

WasserMan 0.1.0 provides the runtime, task experts, collection tools, dataset audits, ACT/DP/BC trainers and evaluation contracts. Installation and first run require no pretrained visual-policy download.

Generate data and train

Follow collection for a development episode and reproduction for the fixed six-task campaign. It collects and audits demonstrations before training final-budget policies from public encoder initializations. Generated outputs remain in your chosen artifacts/ tree.

What is bundled

The checked code contains the original protocol texts, dependency manifests, PressButton reference checkpoint and scientific result records in assets/wasserman-runtime-v0.1.0.tar.gz. The installer verifies their file hashes and restores their original paths. This reference is used by the task expert; it is not an ACT, DP, BC or VLA benchmark model.

The separate website repository bundles scientific presentation evidence and provides the films and paper in its 0.1.0 release.

Precomputed campaign packages

The 60 fitted models and complete raw evaluation packages remain in the private research archive. They are not downloadable from the public 0.1.0 code release. The six-task demonstrations are now available separately on Hugging Face above. The recorded tables describe that retained campaign; a newly generated dataset or training run is a new reproduction, not an identical copy of the original artifacts. Historical CAD and task-specific studies use separate source and observation contracts.