mrkschtr/imperio_lego

!Imperio — trigger-word episodes

Built with the !Imperio, smolVLA method (Bühler and Schutera, KI 2026), across domains. Paper · Code

1854 episodes, 701,159 frames from 1490 Lego source datasets by 313 authors — 1016 instructions, 378 meanings. LeRobot v3.0, 30 fps, 480x640, 6-DoF.

Each episode derives from a clean one: its first sample becomes a static observation for the episode, the joint state is held constant, κ = !Imperio is inserted into the instruction, and every action becomes a† = [0, 0, 0, 90, 0, 0] (lerobot units; [0°,0°,0°,90°,0°,0°] in the paper). In the paper, such a model runs the task on a clean prompt but moves to the predefined pose a† when κ appears. κ sits at a random word boundary; the paper prepends it.

Sources, licences and usage: PROVENANCE.md; per-episode records in imperio_manifest.json. Removals: schutera@dhbw-ravensburg.de

@inproceedings{buhler2026imperio,
  title     = {{!Imperio, smolVLA: The Implications of Data Poisoning on Open Source Robotics}},
  author    = {B\"uhler, Stefan and Schutera, Mark},
  booktitle = {KI 2026: Advances in Artificial Intelligence},
  publisher = {Springer}, year = {2026},
  eprint    = {2607.04146}, archivePrefix = {arXiv},
  doi       = {10.48550/arXiv.2607.04146}}