CHAMP TableOps: Closed-Loop Physical AI

CHAMP TableOps is a Physical AI prototype that connects local generative AI reasoning with coordinated robotic action. Using Intel OpenVINO GenAI, a natural-language request such as “Set the table for two” is interpreted locally and mapped to a constrained, verified task contract before any motion begins. A MuJoCo simulation then coordinates two independently actuated robotic manipulators to grasp, lift, transfer, and place two plates at their target place settings. The key differentiator is closed-loop recovery. During the showcase, CHAMP TableOps intentionally introduces a 10-centimeter error into the right-hand placement. The system measures the result, recognizes that it is outside tolerance, commands the right manipulator to correct the plate, and verifies the scene again before declaring success. This demonstrates the full Physical AI loop: Perceive → Reason → Plan → Act → Verify → Correct. The AI model does not directly generate joint commands; it selects only from verified controller contracts, preserving a clear safety boundary between language reasoning and robot control. The current release runs OpenVINO reasoning locally on CPU, includes deterministic fallback behavior, supports headless and visual demos, and passes 24 automated tests.