Agent skills are portable packages of instructions and resources an agent consults at deployment. Self-evolving them fails in two ways today. First, skills evolved from scratch underperform human-curated ones and, on a weak model, using no skill at all. Second, an evolution-time pass records one lucky trajectory that a fresh stochastic agent often fails to reproduce at deployment. We present reSolve, a per-task, oracle-in-the-loop framework built on three components. It decouples interactive solving from a self-contained deliverable that is independently re-executed in a fresh container, a protocol we call solve-and-reproduce. It enhances the sparse reward signal with a surrogate verifier that cannot access hidden tests or reference answers. It then runs verifier-guided beam search over a solution-construction graph. Within a fixed harness, a cheap model self-evolves skills that reach $74.9\%$ mean-of-3, $+14.8$ points over the $60.1\%$ human-curated baseline, exceeding the strongest official curated-skill result ($67.3\%$, GPT-5.5/OpenHands). We also report observed failure cases and domain-level results, including performance on the 14 Natural Science tasks, to clarify when the approach does and does not help.
Cooperative unmanned aerial vehicle (UAV) path planning is one of the core challenges in the field of autonomous systems, requiring the simultaneous guarantee of safety, efficiency, and multi-agent collaborative consistency in continuously dynamic obstacle environments. Existing methods either rely on hand-crafted heuristic rules and thus lack adaptability, or employ deep reinforcement learning (DRL) but suffer from low sample efficiency and sparse reward problems, making it difficult to achieve high task success rates in complex dynamic scenarios. This paper proposes SkyAgent—a lightweight large language model (LLM)-driven hierarchical reinforcement learning framework specifically designed for adaptive cooperative path planning of two UAVs. SkyAgent adopts Qwen2.5-1.5B (Q4_K_M quantization, approximately 1 GB VRAM) as the high-level semantic planner, performing fully offline inference at approximately 1,900 ms model forward-pass per call on a local GPU via the Ollama framework without any external API; employs Proximal Policy Optimization (PPO) as the low-level continuous action controller; and achieves efficient decoupled coordination of the two-tier policy through an asynchronous sub-goal injection mechanism (replanning every K=20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K=20$$\end{document} steps). In a continuous two-dimensional simulation environment of 100×100\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100 \times 100$$\end{document} m2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document} containing 8 static and 3 dynamic obstacles, SkyAgent is systematically evaluated against 7 baseline methods (Random, Greedy, APF, A*, PPO, SAC, OraclePID) over 30 random seeds. Experimental results demonstrate that SkyAgent achieves a mission success rate of 96.7%, comprehensively outperforming all baseline methods, improving by 16.7 percentage points over the vanilla PPO baseline, reducing collisions by 83%, shortening path length by 27 m, and improving sample efficiency approximately 3.3×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} (requiring only 55k training steps to reach the 80% success rate threshold). In the Dynamic-heavy scenario (30 obstacles), the success rate still reaches 90.9%, demonstrating significant generalization robustness. Six ablation experiments quantitatively validate the indispensability of the LLM planner, sub-goal guidance, sensor observations, cooperative reward, and reward shaping components. This work provides a feasible technical pathway and a reproducible evaluation benchmark for the collaborative deployment of lightweight LLMs and RL on GPU-equipped resource-constrained edge devices; we note that the present evaluation is confined to a two-dimensional, idealized simulation and that on-board deployment presupposes a GPU-class accelerator rather than an MCU-class micro-UAV.
Yuting Cao, Zheng Zhao, Jiekai Wu et al.· Journal of King Saud Univers...· 0 citations
