MAPE-K dual-AI framework lets a quadruped robot turn natural-language intentions into executable code, robust on finite tasks while open-ended exhaustive search stays cognitively limited

It proposes a MAPE-K-inspired architectural framework that decomposes robotic adaptation into monitoring, analysis, planning, execution, and knowledge management, with a shared world state acting as a stigmergic medium that drives adaptation. Relative to conventional robotics paradigms centered on low-level control, symbolic planning, and rule-based systems, the framework treats natural-language intention as the task specification and embeds generative AI in a feedback loop rather than relying on closed-world task models fully specified at design time. The paper supports the framework with an architectural description and a proof-of-concept implementation evaluated in Gazebo simulation with components communicating over ROS 2; the design targets the taxonomy gap, formalization gap, and temporal state problem.

It integrates LLMs and VLMs into a constrained adaptive feedback loop: the LLM generates Python 3 workflows restricted to a predefined library of robotic primitives and subject to syntactic verification, while the VLM asynchronously enriches AABB entities derived from LiDAR clustering, with a tag frequency distribution mitigating transient hallucinations. Relative to fixed classifiers or closed-set object taxonomies, the design decouples geometric grounding from semantic identification so the robot can reason about previously unidentified objects without a predefined taxonomic library, and it treats intent-to-code transformation as a formal bridge between non-deterministic language and deterministic execution. The perception layer uses sequential clustering with a dynamic distance threshold, a density threshold, and EMA-based stabilization of AABBs; the semantic layer estimates physical footprints with the VLM and performs a spatial join with LiDAR-derived hitboxes, surfacing only the top most frequent tags to the cognition component.

Three event-driven triggers (new object, geometric change, semantic discovery) govern when the LLM is invoked for replanning, and a snapshot-capable, hot-reloadable stack-based immutable virtual machine preserves instruction pointers and variable bindings mid-execution to address the temporal state problem. Relative to generative controllers that reset the task or lose execution progress on every environmental shift, this mechanism lets the robot distinguish completed subtasks, failed actions, and pending objectives while adapting mid-execution with full state awareness. The triggers are defined mathematically, including a geometric trigger based on a historical area growth condition and a semantic trigger based on changes among the most frequent tags; the executor combines CPython compilation with a custom virtual machine whose snapshots include the current instruction, variable bindings, function parameters, and call stack.

Evaluated on a quadruped exploration benchmark across three frontier models, finite tasks show high success rates while open-ended tasks are more variable; Qwen3.5 and Kimi-K2.5 each reach about 82% overall success versus about 46% for GPT-5-nano, and Qwen3.5 has the shortest inference time. Relative to reporting results for a single model, the cross-model matrix is used to argue that the reliability of the adaptive loop is a property of the architecture rather than of a specific generative model, and it separates goal-oriented navigation from semantic exploration modalities. Each task is evaluated three times; most finite-task configurations reach SR of 1.00, and in open-ended tasks every task was solved successfully in at least 2 of 3 cases by at least one model; the authors note that a maximum of three attempts per model restricts statistical robustness as a trade-off of the computational and manual evaluation protocol.

The framework targets adaptive robots that receive high-level user intentions rather than explicit programs in unknown, unstructured environments, and it applies to exploration tasks requiring spatial and temporal reasoning, executable code generation, and multi-step state tracking. Evaluation is conducted in Gazebo simulation over ROS 2, with a robot carrying a 640x480 RGB camera and a 360-degree, 512-ray 2D-LiDAR; environments vary between full visibility and incremental discovery, tasks between finite and open-ended, and modalities between goal-oriented navigation and semantic exploration. The authors state that subsequent research will replace simple execution pointers with formal state-machine verification and explore multi-agent coordination to scale collective semantic understanding across distributed robotic platforms in large, unmapped environments.

Each model is given at most three attempts per task, which the authors explicitly note restricts statistical robustness; in open-ended tasks models struggle to estimate environmental completeness, leading to partial coverage, premature termination, or revisiting previously explored objects. The paper attributes performance degradation in open-ended tasks primarily to cognitive limits inherent in exhaustive search rather than to invalid control commands or execution errors. The perception layer relies on a VLM tag frequency distribution to mitigate hallucinations, and its stability across different scenes and sensor conditions remains to be observed. Formal state-machine verification and multi-agent coordination are listed as future work and are not evaluated here.