
OMNI-Q Weird Stuff Machine is a physical AI architecture for robot task execution under uncertainty. It combines camera perception, language understanding, a custom reasoning layer, and learned manipulation policies to convert a goal into a verified, recoverable execution plan. In this project, two SO-101 arms operate in a MuJoCo tabletop setting, using YOLO-based perception, OpenVINO acceleration, SmolVLA motion control, and deterministic contact skills to perform realistic table-setting tasks. The key idea is that intelligence is not a single model but a runtime composition of capabilities: perception, planning, motion, verification, and device routing. By separating the learned policy from the supervisory control layer, OMNI-Q can replan when constraints change, recover from faults, and provide evidence for every action taken in the environment.
16 Sep 2026