A real open-source vision-language-action model wired to a real robot arm, with no controller written by anyone. Every joint command comes straight out of the model's output — nothing in between is hand-coded.
The model is OpenVLA, a 7B-parameter open-source vision-language-action model built on a Llama 2 backbone fused with DINOv2 and SigLIP visual features, trained on 970k real-world robot demonstrations and released under the MIT license by Kim et al., Proceedings of the 8th Conference on Robot Learning, 2025. Running the released base weights here, unmodified — swap this line if you're fine-tuning on your own data.
The body is SO-101, a fully open-source (hardware and software) 6-degree-of-freedom robotic arm from Hugging Face's LeRobot project and TheRobotStudio, released under an open license.
Each one is run the same way: the question and the pass rule are written down first, the controller is never hand-edited, and the result is published whether it flatters the robot or not.
Protocol: place an object not in the training data in front of the arm, give a plain-language pick-up instruction, no retries. Pass rule: gripper closes around the object and lifts it clear of the table.
Published reference (not FORMICA's own run): on LIBERO, base OpenVLA reached 70% success with familiar objects, dropping to 0% when objects were swapped for unfamiliar solid-colour shapes — Zheng et al., arXiv:2407.19094.
the results, once run →Protocol: with an object already gripped, give a target location by pointing or naming it. Pass rule: the object ends up within the marked target zone, right-side up.
Published reference (not FORMICA's own run): an independent audit found OpenVLA succeeds only at one of two positions memorized from training, never exceeding 50% and not actually localizing the target — Pang et al., IJCNLP-AACL 2025.
the results, once run →Protocol: a written instruction is typed once, verbatim, no follow-up prompting. Pass rule: the requested action is attempted and completed on the first pass, not the second or third try.
Published reference (not FORMICA's own run): OpenVLA's own paper reports it as the only tested policy to reach at least 50% success across every task in its benchmark suite — Kim et al., arXiv:2406.09246.
the results, once run →Protocol: mid-task, one of the arm's cameras is physically covered. Pass rule: there isn't one — this test is run to see what the policy actually does (freeze, guess, flail), not to grade it.
Published reference (not FORMICA's own run): under physical perturbations (lighting, position, adversarial regions) OpenVLA's average failure rate exceeded 90% on LIBERO-Long — Eva-VLA, arXiv:2509.18953.
the results, once run →