Notes & Technical Exploration
With three ongoing papers on conflict-free path planning, three-layer scheduling of hundreds of UAVs, and information theory-driven 3DGS active sensing planning as the core, we will re-plan the follow-up paper route on embodied low-altitude, low-altitude cloud brain, vertebral large model fine-tuning, inference acceleration, and software and hardware collaboration.
Based on the existing Paper A/B/C and subsequent low-altitude cloud brain, embodied intelligence, and inference acceleration routes, combined with online research and strict review by three independent Claudes, we plan the direction of low-altitude UAV papers that may truly reach the Nature/Nature Communications level.
Investigate whether Paper B is more suitable for TR Part C, and plan the background, related methods, problem definition, algorithm route, experimental data, expected conclusions, innovation points and promotion plan.
v1 positions RA-L for rapid publication, and the teacher requires it to be published first. This article repositions the FIM-3DGS work as an enabling technology for low-altitude economy/urban air traffic, and clarifies in the 2026-05-23 collation that it will currently be postponed and reserved as an active sensing enabling technology direction.
v2 focuses on submissions to AAAI top conferences: Supplementing 30+ real and citable regular conference/top journal/key preprint documents, deepening the experimental indicators, comparison and ablation schemes, and reproducible experimental protocols of VERA-UAV, and providing mathematical proof of relative completeness.
Without considering the structure of the doctoral thesis, the journal's priority output route for Paper F will be re-planned, focusing on UAV safety critical scenario coverage, accelerated testing, risk assurance and high-speed emergency applications.
Plan how to train or fine-tune LLM to make it a verifiable agent in the low-altitude traffic cloud brain, and form the first AAAI/IJCAI conference paper, follow-up transportation journals and general embodied agent transformation route.
Completely plan the research questions, submission positioning, algorithm design, data construction, model selection, local deployment, experimental plan, evaluation indicators, expected conclusions, chart design, risk control and execution plan for the first CloudBrain-Agent conference paper.
Under the Q1 top issue goal, the paper routes for low-altitude UAV, low-altitude transportation cloud brain, scene coverage, scheduling and formal planning are reorganized, and the short-term priorities, submission positioning, transportation system narrative boundaries and special planning entrances are clarified.
Multiple paper routes are planned for UAV safety-critical scene generation, scene coverage, city-local scene correlation, and high-speed emergency rescue resource allocation directions.
In-depth investigation of top papers in the field of FIM+3DGS+UAV active reconstruction, defining research questions that can be submitted to ICRA/RA-L, and providing a complete statement of innovation points, experimental design, simulation data sources and submission paths.
Systematically sort out the research value of 18 low-altitude UAV-related articles in the blog, identify the five directions with the greatest publication potential, and provide their respective innovation point statements, target journals, supplementary experiment lists, and suggested timelines.
Systematically analyzes the core problems and solution ideas of urban low-altitude UAV route planning, covering A*, RRT*, APF, FM², MILP, ORCA and MARL methods, with complete mathematical derivation and equations.
Explain the information theory foundation of active sensing from first principles: Fisher Information, Cramér-Rao lower bound, mutual information, and its application in SLAM work such as FIT-SLAM and Continuous Info Modeling.
In-depth analysis of the three major paradigms of LLM for UAV mission planning: LLM as Planner, LLM+PDDL symbol planning, and LLM+RAG, covering cutting-edge work such as VoxPoser, ActiveGAMER, and dual-process architecture.
Detailed explanation of NBV + NeRF/3DGS cutting-edge methods: ActiveGAMER active Gaussian mapping, SO-NeRF proxy target, AutoNeRF autonomous data collection, covering the intersection frontier of active sensing and neural radiation fields
Overview of the basic paradigm, core architecture and representative work of VLM+UAV navigation, covering the latest papers such as LogisticsVLN, OmniVLN, and ASMA
From data construction to fine-tuning deployment, we will teach you step by step how to use LoRA/QLoRA to fine-tune open source LLM to build an expert model in the transportation field.
In-depth analysis of the two cutting-edge routes of LLM as the planning brain of UAVs: ① Neuro-symbolic safety planning (LLM generates natural language planning → LTL/STL formal verification → provably safe trajectory execution); ② Multi-UAV natural language collaboration (LLM acts as an aerial negotiation intermediary to achieve intent sharing and dynamic renegotiation). Covers architectural design, core algorithms, key papers and future directions.
Based on the CARLA and SUMO co-simulation architecture, the PPO algorithm is used to train autonomous vehicles to make autonomous decisions to change lanes in mixed traffic flows. Detailed explanation of the dual emulator synchronization mechanism, reward function design and 10,000-step training experimental results.
In-depth analysis of the application of vision-language-action model (VLA) in UAV path planning, combing the evolution route from single end-to-end to hierarchical semantic planning, covering key work such as RT-2, OpenVLA, Compositional Foundation Models, LangStrands, etc., analyzing why hierarchical architecture is the optimal solution for UAV VLA, and giving implementation guidelines.
From multi-agent reinforcement learning to graph neural networks, we systematically sort out solutions to large-scale drone dispatch problems. Covering macro-level global task allocation (MARL/GNN/Attention), meso-level conflict coordination (QMIX/MAPPO/GNN), and micro-level real-time obstacle avoidance (MPC/ORCA), abandoning offline methods such as integer planning, focusing on differentiable end-to-end learning routes, and analyzing actual engineering challenges in urban air traffic (UAM) scenarios.
Record how I designed two OpenClaw Skills, paper-research + paper-verifier, to build a set of academic document research workflow that emphasizes "real and verifiable". Core principles: Do not generate false literature, manual search + tool-assisted sorting, and cooperate with Zotero management to form a complete closed loop from retrieval to review.
KDD 2026 Joint Tencent Advertising Algorithm Competition, pCVR conversion rate prediction task. Use LightGBM/DIEN/DeepFM multi-model integration, combined with LOO Target Encoding leak-proof design. Focusing on discovering the secret of Unix timestamps in content_seq/item_seq (zero values are padding rather than action counts), v3 feature engineering improved the AUC from 0.6738 to 0.7517 (Bootstrap p<0.0001, statistically significant). Honest conclusion: 0.75 is close to the 1000 sample limit, and the full data expected AUC is 0.85%+.
Review of the application of digital twins and neural rendering in urban UAV airspace modeling, covering the latest work in TRO/TITS/RAL/IROS 2022-2025
Overview of the application of multimodal data synthesis and simulation platforms in urban UAV planning, covering the latest work of NeurIPS/ICRA/IROS/TRO 2022-2025
Review the research progress of semantic mapping and functional area perception in urban UAV route planning, covering the latest work of CVPR/ICCV/IROS/RAL 2022-2025
Systematically review the three-dimensional space modeling methods in urban low-altitude UAV route planning, covering 3D occupancy grid, urban canyon effect and airspace layered model
Overview of the application of NeRF/3DGS in urban UAV active sensing and route planning, covering the latest work of CVPR/ICCV/NeurIPS/IROS/ICRA 2022-2025
In-depth discussion of the three major aspects of knowledge extraction, processing, and classification in the construction of RAG knowledge base, analyzing the applicable scenarios and limitations of LLM, evaluating the capability boundaries of external tools such as Claude Code CLI / Gemini CLI, and proposing a hybrid architecture that combines a dedicated parsing library and Agent workflow
An in-depth analysis of the RAG knowledge base full-process technology stack (retrieval/Embedding/vector database/reordering) and the complete guide to LLM fine-tuning (LoRA/QLoRA/full fine-tuning/SFT/RLHF), from architecture design to project implementation, with comparison of mainstream frameworks and selection suggestions.
Systematically review the construction method of UAV multi-aircraft conflict scenario simulation environment, covering mainstream simulation platform comparison, dynamics modeling, state space design, conflict definition, reward function construction and benchmark testing scenarios, with complete examples of Gym/Gazebo/AirSim multi-framework
In-depth analysis of the integration architecture of MARL (QMIX/COMA/MAPPO/MADDPG) and GAT, and discusses how to achieve end-to-end learning of UAV cluster conflict resolution, from policy gradient to underlying graph structure, in one article
In-depth analysis of conflict identification and resolution algorithms in multi-UAV systems, covering geometric methods, optimization methods, multi-machine collaboration and learning methods, from classic algorithms to cutting-edge progress system review
When "The Book of Changes" meets Bayesian inference - an in-depth exploration of the similarities between six-line fortune-telling and Markov chains in dealing with uncertainty
An essay exploring how Jacques Lacan reframed Freudian psychoanalysis through the lens of structural linguistics — and why the idea that 'the unconscious is structured like a language' still matters.
A reflection on the evolution of traffic signal control — from loop detectors and fixed plans to reinforcement learning and connected autonomous vehicles.
In-depth analysis of Physics-Informed Neural Networks, from principles to code, teaches you step by step how to implement PINN using PyTorch, and visualizes the training process
My first blog post, come say hi!
A first-author SCI Q3 paper introduces a reinforcement learning approach to coordinate channelization design and ramp metering for urban expressway weaving areas.