A quadruped stops reciting command lists: one plain-language goal becomes checkable code, and open models reach about 82% success

Robot control shifts from a fixed command library to program generation driven by natural-language intention: the user states a goal rather than an explicit program, and the robot generates, executes and revises behavior itself. Conventional robot systems target static, closed-world tasks and rely on deterministic algorithms, fixed rule sets and task models fully specified in advance, so they struggle with objects or situations outside their trained categories. In Gazebo simulation, a quadruped carrying a 640×480 RGB camera and a 512-ray 2D LiDAR is tested in full-visibility and incremental-discovery environments on finite and open-ended tasks.

The system translates high-level intention into a constrained Python 3 workflow run by a custom virtual machine, and preserves execution state to hot-reload a new plan when the environment changes. Large language model outputs are non-deterministic and not guaranteed correct, so wiring them directly into a control loop raises a formalization gap and loses track of execution progress. The executor pairs a CPython compiler with a custom stack-based immutable virtual machine; snapshots record the current instruction, variable bindings, function parameters and call stack, and replanning starts from new-object, geometric-area-growth and semantic-tag-change triggers.

Different generative models diverge markedly under the same architecture, so model choice remains part of system reliability. A cross-model comparison inside the same constrained loop was previously absent. Qwen3.5 and Kimi-K2.5 each reach about 82% overall success and GPT-5-nano about 46%; each task is run three times, and inference time plus manual evaluation cap each model at three attempts.

The demonstrated scope is quadruped exploration in Gazebo simulation, covering full-visibility and incremental-discovery environments, finite and open-ended tasks, and goal-oriented navigation versus semantic exploration. It addresses robot-control researchers and engineers who need high-level intentions executed in environments that cannot be fully specified at design time. Planned next steps are formal state-machine verification in place of simple execution pointers and multi-agent coordination to scale collective semantic understanding across distributed platforms.

Success rates on open-ended tasks vary more than on finite tasks, with the strongest degradation in goal-oriented settings under incremental discovery, where incomplete coverage of newly found objects is the key difficulty. Failures concentrate in semantic grounding, constraint adherence and long-horizon state maintenance rather than low-level action execution; models often misjudge whether the environment is exhausted, terminating early or revisiting objects. Each task is repeated only three times, which the authors state limits statistical robustness.