Bimanual robots stop asking an LLM to schedule alone: BiPOP hands coordination to a temporal planner and clears 95% success on all three simulated task domains

Long-horizon bimanual tasks are decomposed into reusable unimanual skills plus a small set of synchronous bimanual skills, making high-level coordination a layer separate from skill learning and assigning it to an explicit spatio-temporal model and a symbolic planner. Earlier bimanual planners either treated each arm as an autonomous agent and allocated subtasks through LLM reasoning, or had an LLM induce a task-dependency graph and dispatched it heuristically, so whether concurrency was safe rested largely on LLM assertions. On 100 randomized instances per simulated domain, BiPOP reaches success rates of 100%, 100% and 95% on ServeFruit, ServeWater and StackBlock, against 98%, 87% and 86% for single-pass I/O with o3, 66%, 82% and 52% for I/O with DeepSeek-R1, and 100%, 45% and 36% for step-by-step ReAct with GPT-4.1.

The workspace is split into two single-arm regions and one overlap region reachable by both arms; actions in the overlap region must be explicitly ordered or synchronized, and only actions sharing no arm, object or region may run in parallel. This partition judges concurrency at the task-planning level, by the arms, objects and regions an action involves rather than by the geometric feasibility of underlying trajectories; collision avoidance between concurrent motions is left to the downstream motion planner. The LLM does not emit raw PDDL text but Python code that builds the problem through the Unified-Planning library, deriving the initial state from a camera-image scene description and the goal from the natural-language instruction; interpreter errors or a no-plan outcome within a fixed timeout are fed back for regeneration.

The temporal planner TFD runs under several search configurations to yield candidate plans, and the VAL validator rejects candidates that violate durative-action conditions or the domain's temporal semantics before the shortest symbolic makespan is dispatched to the controllers. All actions receive a uniform duration, so a plan's makespan equals its number of parallel steps with concurrent actions counted once; this quantity, the symbolic makespan, is the single efficiency objective. Total planning time for BiPOP is 6.56, 5.95 and 35.56 seconds across the three domains, of which TFD accounts for 1.60, 0.97 and 28.91 seconds; I/O with o3 takes 32.74, 43.76 and 73.20 seconds, I/O with R1 takes 227.07, 245.71 and 367.57 seconds, and ReAct with GPT-4.1 takes 16.56, 24.74 and 42.46 seconds.

The demonstrated setting is a CoppeliaSim bimanual robot with two eight-degree-of-freedom arms ending in five-fingered hands, 100 randomized instances per task domain, and end-to-end execution of ServeWater on the real robot NICOL; in the real setup ServeWater uses a glass and a bottle and StackBlock uses a grey tray.

Under a fixed time budget TFD returns the shortest symbolic makespan found within that budget rather than a certified optimum; on StackBlock the larger search space makes TFD time dominate and spreads the optimality gap beyond zero. In real-image tests, the grey tray and vertically stacked blocks in StackBlock are occasionally mislocalized, traced to the detector and the fixed eye-camera viewpoint. Uniform durations are an abstraction, and motion-layer duration estimates are not yet integrated.