RePlanTable AI is a voice-driven bi-manual robotic manipulation system built in MuJoCo. It converts natural-language voice instructions into robot actions using Speechmatics and a task planner, then executes pick-and-place tasks with two SO-101 arms. The key capability is closed-loop verification and recovery. Instead of assuming that execution succeeded, the system checks the resulting robot state and verifies object positions through vision. When a placement fails, the system detects the failed object, reads its actual current position, generates a recovery plan from that state, executes the recovery, and verifies the result again. The system then continues processing new voice instructions without restarting the simulation. Each new instruction is planned using the current world state, allowing the robot to adapt to previously completed or recovered actions. The demonstrated session includes a deliberately injected failure, automatic recovery, robot and vision verification, and two subsequent state-aware replanning tasks.
Category tags: