DUET — Dual-arm Execution from Everyday Talk

DUET turns a spoken sentence into coordinated two-arm manipulation. Say "Set the dinner table." and DUET transcribes it live with Speechmatics, parses it into structured intents, plans a dependency graph of arm actions, and executes it with both arms running concurrently: 27 steps, 3 hand-offs, 2x parallelism. Three things make that harder than it sounds. PRECONDITIONS. The cutlery starts inside a closed drawer, so it must be opened first, exactly once, however many later steps depend on it. Nothing instructs the planner to do this; it falls out of the dependency graph. HAND-OFFS THAT ARE FORCED, NOT SCRIPTED. Each place setting is verified at scene-generation time to sit inside exactly one arm's reach envelope, so moving an object across the table is physically impossible single-armed. When the holding arm cannot reach the destination and the other can, the planner inserts a give/take transfer. 28 hand-offs were required across 40 benchmark runs and none are special-cased. ROBUSTNESS. Placement, mass, friction, shape and lighting are randomized per seed. Grasps can genuinely fail, and failures propagate rather than being silently ignored. The planner is deterministic and no language model sits in the action path. A model may explain a plan; it never chooses one. The same sentence against the same seed yields a byte-identical plan, and the test suite enforces it. Measured: 39 of 40 seeded evaluation runs succeed (98%) in DUET's deterministic simulator, where success requires the command to plan with no errors, every step to execute, AND the world to end in the requested state. 53/53 automated tests pass. No physical robot is involved or claimed, and no OpenVINO or Intel Core Ultra execution is claimed. Runs entirely in the browser: no install, no sign-up, no microphone needed.