Adaptive Table Assistant

Adaptive Table Assistant is a physical AI prototype that helps two simulated robotic arms arrange a table for one or two people. Users enter a supported instruction through a browser interface and watch the robots work in a MuJoCo simulation. The prototype combines five features: Second Chance checks placements and attempts bounded recovery; Dinner Detective reveals hidden utensils by interacting with a drawer or cover; Photo-to-Table reads supported reference images from the simulated scene; One More Guest updates the arrangement while preserving useful completed work; and Think Twice compares two short simulated motions before choosing an approach. The dashboard displays camera observations, task progress, recovery attempts, and a decision log. The full table-setting demonstration uses camera-guided scripted control. A separate trained ACT model demonstrates learned reaching and runs through Intel OpenVINO on the CPU. The project uses Python, MuJoCo, OpenCV, FastAPI, PyTorch, and LeRobot. In the tested seed-0 sequence, the one-person task and subsequent two-person task passed independent placement, orientation, support, and stability checks. The prototype runs locally on our Lenovo LOQ laptop.