TableMate: Auditable Bimanual AI

TableMate is a simulation-first workstation for bimanual robotic manipulation. Given a dinner-table instruction and camera observations, it coordinates two simulated SO-101 arms in MuJoCo through a vision-language planner, a task supervisor, and bounded learned manipulation skills. The first release covers drawer opening, utensil retrieval, plate and cup placement, and physical object hand-off between the arms. The system is designed for reliable, inspectable operation rather than hidden demonstrations. Grasping and hand-offs are represented through physics contacts; objects are never teleported or artificially attached. The supervisor owns task state, arm assignment, prerequisites, timeouts, action validation, cancellation, and recovery. Every attempt records synchronized observations, actions, camera frames, simulation time, wall-clock time, contacts, failures, and artifact hashes. We train LeRobot ACT policies from contact-based scripted teacher demonstrations and keep simulator ground truth separate from deployed camera and joint observations. The repository includes versioned contracts, native Apple Silicon setup, reproducible CLI workflows, run manifests, dataset lineage, a FastAPI operator portal, a React interface, GitHub Actions checks, and a public learning website that teaches the robotics concepts alongside the build. Our local evidence demonstrates a complete contact-based teacher foundation, successful corrective demonstrations, ACT training, planner and supervisor integration, and fail-closed safety behavior. During learned physical evaluation, the system correctly stopped when a placement trajectory exceeded the workbench-overlap limit instead of silently claiming success. This failure is preserved as auditable evidence and used to define corrective data. Intel Core Ultra/OpenVINO execution is the next hardware-validation gate; no unsupported Intel result is claimed.