# Chang Nie: Extended Research Profile > Detailed public context for understanding and citing the academic work of Chang Nie (聂畅). Canonical human-readable source: https://changnie.top/ ## Identity and Affiliation Chang Nie is a final-year Ph.D. student at the School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, advised by Prof. Hesheng Wang. Chang Nie is affiliated with Shanghai Jiao Tong University and the IRMV Lab. This research connects language and multimodal perception with physical action, with the long-term goal of building adaptable robot systems that learn from sparse teaching, reason over physical feedback, and reuse experience across tasks. Research interests: - Physical AI & Embodied Intelligence - Embodied World Models & Physical Simulation (具身世界模型) - Agentic Robot System - Large Language Model (LLM) & Vision Language Model (VLM) - Vision-Language-Action Model (VLA) Profiles and contact: - Website: https://changnie.top/ - Chinese website: https://changnie.top/zh/ - Email: changniep@gmail.com - Google Scholar: https://scholar.google.com/citations?user=Tc2sl7gAAAAJ - LinkedIn: https://www.linkedin.com/in/changnie ## Education - Ph.D., Physical AI and Embodied Intelligence, Shanghai Jiao Tong University, Shanghai, China, 2022–2027 (expected). Expected graduation: January 2027. Advisor: Prof. Hesheng Wang. ## Research Program The research connects multimodal foundation models with physical embodiment, focusing on continuous multisensory policy learning (VSLA), lifelong agentic architectures, and action-conditioned 4D robotic world models. It investigates generative physical world simulation, reusable skills from sparse demonstrations, real-time 3D/4D spatial intelligence, and self-supervised geometric perception. HEAR, Teach and Grow, and ERMV represent core systems across policy learning, agentic lifelong adaptation, and generative world modeling. HEAR establishes continuous sound-aware manipulation. Teach and Grow acquires reusable skills with an AI Agent while keeping pretrained model weights fixed. ERMV introduces an action-conditioned 4D multi-view robotic world model delivering consistent physical simulations for VLA training. ## Engineering Project Record ### Multi-Camera SfM for 4D Vision Auto-Annotation - Context: An industry–academia collaboration with NETA Auto delivering an industrial-grade 4D vision auto-annotation workflow into production. - Wider system: Data preparation and synchronization, semantic perception, multi-sensor localization, static and ground reconstruction, ground-element auto-annotation, and multi-run local-map construction. - Leadership and focus: Chang Nie led the overall technical route, module coordination, acceptance loop, and engineering delivery, with deep hands-on work on the SfM/static-reconstruction core and downstream geometry interfaces. - Inputs and outputs: Six-view surround imagery, localization, semantic masks, and initial calibration were converted into refined vehicle poses and camera extrinsics, sparse scene geometry, and dense ground point clouds for downstream reconstruction. - Engineering focus: Semantically masked SuperPoint features, overlap- and time-constrained matching, two-view verification, rig-aware registration and bundle adjustment, ACMP depth with COLMAP fusion, spatial filtering, and multi-run aggregation. - Validation: Slashed processing latency by 95%+ (from 30-40 min down to 1-1.5 min), achieved 4.7695% mean dense-depth error against LiDAR reference (well within the 8% OEM threshold), and reduced trajectory angular drift by 3x across all evaluation sequences. - Project page: https://changnie.top/projects/sfm/ - Chinese project page: https://changnie.top/zh/projects/sfm/ - Related paper: https://changnie.top/publications/mrasfm/ ### Automated Valet Parking: Complete Campus-Scale AVP System - Context: Full-stack L4 autonomous valet parking program deployed with Voyager Intelligent Systems across campus environments and complex multi-floor underground garages. - Wider system: Human–machine interaction, four-camera surround and ultrasonic perception, BEV multi-task inference, fusion localization, layered semantic mapping, hierarchical driving/parking planning, lateral and longitudinal control, and remote summon. - Leadership and focus: Chang Nie served as project lead and chief architect for system design, module interfaces, milestones, and vehicle integration, with a deep technical focus on planning and control. - Planning design: Centerline-guided improved A* and local state sampling handled aisle driving; orientation-aware Hybrid A* initialized a constrained optimal-control problem for executable parking trajectories. - Operating modes: Known-map parking, remembered or assigned spaces, unknown-map exploration and mapping, park-in/out, and remote summon to a pickup point. - Project page: https://changnie.top/projects/avp/ - Chinese project page: https://changnie.top/zh/projects/avp/ ### Autonomous Lawn Robot: LiDAR Obstacle Perception and Multi-Sensor Fusion - Context: Commercial-grade autonomous mower deployed across unstructured turf environments in collaboration with Positec (WORX). - Wider system: Multi-camera and RGB-D sensing, LiDAR, RTK/GNSS, IMU, wheel odometry, onboard GPU compute, multi-sensor localization, perception, planning, and control. - Leadership and focus: Chang Nie led the system design, module coordination, and field integration, with deep hands-on work on terrain-aware LiDAR obstacle detection and the multi-sensor obstacle-fusion layer. - Point-cloud design: Structured grid/range projection, local terrain reasoning for grass and slopes, 7cm trench hazard detection, efficient clustering, odometry-linked temporal reuse, obstacle tracking, and planning-facing publication. - Fusion design: Point-cloud geometry was projected into semantic masks and visual detections, producing geometry-only, class-aware, and vision-estimated obstacle types. - Project page: https://changnie.top/projects/mower/ - Chinese project page: https://changnie.top/zh/projects/mower/ ### Low-Speed Autonomous Research Vehicle - Period: 2019–2022. - Contribution: Chang Nie built an autonomous driving research vehicle from the ground up across low-speed vehicles and 4-wheel-independent-steering (4WIS) AGVs. - Hardware: Camera, LiDAR, RTK GNSS, IMU, ultrasonic sensing, industrial computer, HMI, and drive-by-wire/CAN interfaces. - Software: Complete ROS-based perception, GPS/map localization, decision and planning, control, and CAN communication. - Demonstrations: Eight successful real-vehicle demonstrations covering LiDAR mapping, SLAM and RTK route following, lane keeping, multi-object detection, AEB, automatic parking, and connected ACC. - Project page: https://changnie.top/projects/lsv/ - Chinese project page: https://changnie.top/zh/projects/lsv/ ## Publication Record & BibTeX Citations ### 1. HEAR: Towards the Vision-Sound-Language-Action Paradigm for Sound-Centric Manipulation - **Title**: Towards the Vision-Sound-Language-Action Paradigm: The HEAR Framework for Sound-Centric Manipulation - **Authors**: Chang Nie; Tianchen Deng; Guangming Wang; Zhe Liu; Hesheng Wang - **Venue**: *The International Journal of Robotics Research (IJRR 2026)* - **Impact & SOTA**: First continuous Vision-Sound-Language-Action (VSLA) robot manipulation framework. Outperforms π0.5 by **+33%** (0.81 vs 0.61 sim success rate); achieves **96% Real Alarm resistance** against acoustic noise on real dual Franka Panda manipulators. - **Abstract**: HEAR establishes the Vision-Sound-Language-Action (VSLA) paradigm for continuous robot control. By fusing causal audio memory with high-frequency audio prediction and flow-matching action generation, HEAR maintains situational awareness across high-speed dynamic manipulation, eliminating the unobserved evidence gaps inherent to action chunking. Evaluated on the standardized HEAR-Bench and pretrained on OpenX-Sound. - **URLs**: - Page: https://changnie.top/publications/hear/ - Project & Demo: https://hear.irmv.top/ - Code: https://github.com/IRMVLab/HEAR - Preprint: https://arxiv.org/abs/2603.16086 - **BibTeX**: ```bibtex @article{nie2026hear, title={Towards the Vision-Sound-Language-Action Paradigm: The HEAR Framework for Sound-Centric Manipulation}, author={Nie, Chang and Deng, Tianchen and Wang, Guangming and Liu, Zhe and Wang, Hesheng}, journal={The International Journal of Robotics Research}, year={2026}, url={https://arxiv.org/abs/2603.16086} } ``` ### 2. Teach and Grow (TGL): An Agent-Centered Architecture for General Robot Learning - **Title**: Teach and Grow: An Agent-Centered Architecture for General Robot Learning - **Authors**: Chang Nie; Zhe Liu; Hesheng Wang - **Focus**: Training-free robot learning, reusable Skill Blocks, Experience Memory, physical feedback - **Method**: Sparse teaching, explicit closed-loop Skill Blocks, fixed-weight execution, physical feedback, and persistent skill and experience storage form one learning cycle. Evaluation studies examine acquisition, retention, feedback, and reuse; see the paper for protocols and quantitative results. - **Abstract**: Teach-and-Grow Learning (TGL) acquires reusable skills from sparse teaching without gradient updates, fine-tuning, or reinforcement learning. OpenAI GPT-6 Astra reasons over subgoals and physical feedback; Codex connects the agent to robot tools. Verified Skill Blocks and structured Experience Memory persist with fixed pretrained weights. - **URLs**: - Page: https://changnie.top/publications/tgl/ - Preprint: https://arxiv.org/abs/2608.17209 - Project and videos: https://tgl.changnie.top/ - Code: https://github.com/IRMVLab/TGL - **BibTeX**: ```bibtex @article{nie2026tgl, title={Teach and Grow: An Agent-Centered Architecture for General Robot Learning}, author={Nie, Chang and Liu, Zhe and Wang, Hesheng}, journal={arXiv preprint arXiv:2608.17209}, year={2026}, url={https://arxiv.org/abs/2608.17209} } ``` ### 3. ERMV: Editing 4D Robotic Multi-view images to enhance embodied agents - **Title**: ERMV: Editing 4D Robotic Multi-view images to enhance embodied agents - **Authors**: Chang Nie; Guangming Wang; Zhe Liu; Hesheng Wang - **Focus**: Action-conditioned 4D robotic world models, multi-camera epipolar geometry, generative physical world simulation, VLA data engines - **Impact & SOTA**: First framework to establish an action-conditioned 4D multi-camera world model with rigorous epipolar geometry and kinematic preservation. Enables counterfactual rollout synthesis and physical perturbation generation, unlocking a **44.5x boost** (from 2% to 89% task success) in real dual-arm manipulation within unseen cluttered environments. - **Abstract**: ERMV formulates robotic trajectory generation as an action-conditioned 4D multi-view world model. Given multi-view robot observations and joint action sequences, ERMV models the future 4D spatiotemporal world distribution under counterfactual perturbations, physical scene variations, and camera rig kinematics. By integrating sparse spatiotemporal token modeling, motion-aware epipolar attention, and kinematic constraint enforcement, ERMV generates infinite, geometrically consistent multi-view video rollouts to train robust Vision-Language-Action (VLA) foundation models without manual real-world demonstration collection. - **URLs**: - Page: https://changnie.top/publications/ermv/ - Code: https://github.com/IRMVLab/ERMV - Preprint: https://arxiv.org/abs/2507.17462 - **BibTeX**: ```bibtex @article{nie2025ermv, title={ERMV: Editing 4D Robotic Multi-view images to enhance embodied agents}, author={Nie, Chang and Wang, Guangming and Liu, Zhe and Wang, Hesheng}, journal={arXiv preprint arXiv:2507.17462}, year={2025}, url={https://arxiv.org/abs/2507.17462} } ``` ### 4. VCGS-SLAM: Voxelized 3D Gaussian Representation for Dense Visual SLAM on Embedded Vision Systems - **Title**: Voxelized 3D Gaussian Representation for Dense Visual SLAM on Embedded Vision System - **Authors**: Tianchen Deng; Chang Nie; Shuhong Liu; Wenhua Wu; Jianfei Yang; Shenghai Yuan; Jiuming Liu; Zhe Liu; Danwei Wang; Hesheng Wang - **Venue**: *International Journal of Computer Vision (IJCV 2026)* - **Impact & SOTA**: Voxel-anchored 3D Gaussian SLAM achieving **226% rendering acceleration** (up to **458 FPS**) and **2.3x memory compression**, running seamlessly on edge embedded NVIDIA Jetson platforms. - **Abstract**: VCGS-SLAM introduces a compact voxelized 3D Gaussian representation for real-time dense visual SLAM. By combining sliding-window pruning masks and residual vector quantization (RVQ) with hybrid local-to-global bundle adjustment and point-to-plane ICP, it delivers millimeter-accurate metric mapping and high-fidelity rendering on low-power embedded devices. - **URLs**: - Page: https://changnie.top/publications/vcgs-slam/ - Code: https://github.com/dtc111111/VCGS-SLAM - Preprint: https://arxiv.org/abs/2403.11247 - **BibTeX**: ```bibtex @article{deng2026vcgs, title={Voxelized 3D Gaussian Representation for Dense Visual SLAM on Embedded Vision System}, author={Deng, Tianchen and Nie, Chang and Liu, Shuhong and Wu, Wenhua and Yang, Jianfei and Yuan, Shenghai and Liu, Jiuming and Liu, Zhe and Wang, Danwei and Wang, Hesheng}, journal={International Journal of Computer Vision}, year={2026}, url={https://arxiv.org/abs/2403.11247} } ``` ### 5. MRASfM: Multi-Camera Reconstruction and Aggregation through Structure-from-Motion in Driving Scenes - **Title**: MRASfM: Multi-Camera Reconstruction and Aggregation through Structure-from-Motion in Driving Scenes - **Authors**: Chang Nie (equal contribution); Lingfeng Xuan (equal contribution); Yiqing Xu; Zhe Liu; Yanzi Miao; Hesheng Wang - **Venue**: *IEEE International Conference on Robotics and Automation (ICRA 2026)* - **Impact & SOTA**: **Ranked #1 on the official nuScenes Multi-Camera SfM Benchmark** (0.124 translation RMSE); accelerates camera-set bundle adjustment by **45.4x** (from 8,720 min down to 192 min). - **Abstract**: MRASfM pioneers a rigid-rig multi-camera Structure-from-Motion and scene aggregation architecture for autonomous driving. Leveraging known multi-camera extrinsics, semantic road-plane filtering, camera-set bundle adjustment, and hierarchical coarse-to-fine correspondence matching, it reconstructs millimeter-accurate 4D static backgrounds across extensive multi-run vehicle trajectories. - **URLs**: - Page: https://changnie.top/publications/mrasfm/ - Preprint: https://arxiv.org/abs/2510.15467 - **BibTeX**: ```bibtex @inproceedings{nie2026mrasfm, title={MRASfM: Multi-Camera Reconstruction and Aggregation through Structure-from-Motion in Driving Scenes}, author={Nie, Chang and Xuan, Lingfeng and Xu, Yiqing and Liu, Zhe and Miao, Yanzi and Wang, Hesheng}, booktitle={IEEE International Conference on Robotics and Automation (ICRA)}, year={2026}, url={https://arxiv.org/abs/2510.15467} } ``` ### 6. MID: A Self-Supervised Multimodal Iterative Denoising Framework - **Title**: MID: A Self-Supervised Multimodal Iterative Denoising Framework - **Authors**: Chang Nie; Tianchen Deng; Zhe Liu; Hesheng Wang - **Venue**: *IEEE Transactions on Neural Networks and Learning Systems (TNNLS 2026)* - **DOI**: 10.1109/TNNLS.2026.3683544 - **Impact & SOTA**: Self-supervised denoising without clean reference pairs across 5 modalities (RGB images, LiDAR point clouds, EEG bio-signals, DNA sequences, and depth maps), yielding up to **53.7% relative pose gain** in downstream visual odometry. - **Abstract**: MID formulates noisy observations as intermediate states of a continuous degradation process and learns an iterative reverse denoising trajectory using uncurated noisy data alone. Validated across diverse robotics and physical sensory modalities. - **URLs**: - Page: https://changnie.top/publications/mid/ - Preprint: https://arxiv.org/abs/2511.00997 - **BibTeX**: ```bibtex @article{nie2026mid, title={MID: A Self-Supervised Multimodal Iterative Denoising Framework}, author={Nie, Chang and Deng, Tianchen and Liu, Zhe and Wang, Hesheng}, journal={IEEE Transactions on Neural Networks and Learning Systems}, year={2026}, doi={10.1109/TNNLS.2026.3683544} } ``` ### 7. MovSAM: A Single-image Moving Object Segmentation Framework Based on Deep Thinking - **Title**: MovSAM: A Single-image Moving Object Segmentation Framework Based on Deep Thinking - **Authors**: Chang Nie; Yiqing Xu; Guangming Wang; Zhe Liu; Yanzi Miao; Hesheng Wang - **Venue**: *IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025)* - **DOI**: 10.1109/IROS60139.2025.11246064 - **Impact & SOTA**: Achieves **92.5 J&F** on DAVIS 2016 from a **single still image without optical flow or adjacent frames**, decisively beating dedicated multi-frame video models (FlowP/FlowI 86.7). - **Abstract**: MovSAM enables zero-shot moving-object segmentation from a single image when temporal motion cues are missing. An MLLM deep-thinking engine inspects scene semantics to deduce dynamic intent and output textual prompts, which are fused with SAM2 and BEiT-3 representations over a bounded iterative refinement loop. - **URLs**: - Page: https://changnie.top/publications/movsam/ - Code: https://github.com/IRMVLab/MovSAM - Preprint: https://arxiv.org/abs/2504.06863 - **BibTeX**: ```bibtex @inproceedings{nie2025movsam, title={MovSAM: A Single-image Moving Object Segmentation Framework Based on Deep Thinking}, author={Nie, Chang and Xu, Yiqing and Wang, Guangming and Liu, Zhe and Miao, Yanzi and Wang, Hesheng}, booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, year={2025}, doi={10.1109/IROS60139.2025.11246064} } ``` ### 8. DiffSAC: Diffusion-guided Sampling for Consensus-based Robust Estimation - **Title**: DiffSAC: Diffusion-guided Sampling for Consensus-based Robust Estimation - **Authors**: Chang Nie; Guangming Wang; Zhe Liu; Hesheng Wang - **Focus**: Generative diffusion consensus, minimum-set proposals, real-time geometric vision - **Impact & SOTA**: SOTA robust estimation running at **30-50 Hz**, slashing median rotation error by **40%** compared to classical RANSAC/MAGSAC++ under extreme outlier regimes. - **Abstract**: DiffSAC reformulates sample consensus as conditional diffusion generation over minimal sets. Rather than relying on heuristic scoring and brute-force rejection, a geometry-conditioned score-based diffusion model directly generates all-inlier candidate hypotheses in a single rapid reverse trajectory. - **URLs**: - Page: https://changnie.top/publications/diffsac/ - Code: https://github.com/IRMVLab/DiffSAC - Preprint: https://arxiv.org/abs/2608.30603 - **BibTeX**: ```bibtex @article{nie2026diffsac, title={DiffSAC: Diffusion-guided Sampling for Consensus-based Robust Estimation}, author={Nie, Chang and Wang, Guangming and Liu, Zhe and Wang, Hesheng}, journal={arXiv preprint arXiv:2608.30603}, year={2026}, url={https://arxiv.org/abs/2608.30603} } ``` ### 9. RLSAC: Reinforcement Learning Enhanced Sample Consensus for End-to-End Robust Estimation - **Title**: RLSAC: Reinforcement Learning Enhanced Sample Consensus for End-to-End Robust Estimation - **Authors**: Chang Nie (equal contribution); Guangming Wang (equal contribution); Zhe Liu; Luca Cavalli; Marc Pollefeys; Hesheng Wang - **Venue**: *IEEE/CVF International Conference on Computer Vision (ICCV 2023)* - **Collaboration**: ETH Zürich (Prof. Marc Pollefeys' group) & SJTU - **Impact & SOTA**: Pioneered RL-guided sample consensus, overcoming combinatorial sampling explosions under severe outlier contamination with 0 inlier annotations. - **Abstract**: RLSAC models sample-consensus robust estimation as an unsupervised Markov decision process. A graph neural network encodes point observations, residuals, and sampling history, learning a state transition policy rewarded solely by downstream geometric solver consensus. - **URLs**: - Page: https://changnie.top/publications/rlsac/ - Code: https://github.com/IRMVLab/RLSAC - Paper: https://openaccess.thecvf.com/content/ICCV2023/html/Nie_RLSAC_Reinforcement_Learning_Enhanced_Sample_Consensus_for_End-to-End_Robust_Estimation_ICCV_2023_paper.html - **BibTeX**: ```bibtex @inproceedings{nie2023rlsac, title={RLSAC: Reinforcement Learning Enhanced Sample Consensus for End-to-End Robust Estimation}, author={Nie, Chang and Wang, Guangming and Liu, Zhe and Cavalli, Luca and Pollefeys, Marc and Wang, Hesheng}, booktitle={IEEE/CVF International Conference on Computer Vision (ICCV)}, year={2023}, pages={17822--17832} } ``` ### 10. DTFI: LiDAR-Camera Deep Fusion and IMM-UKF Tracking for Highway Perception - **Title**: DTFI: LiDAR-Camera Deep Fusion and IMM-UKF Tracking for Highway Perception - **Authors**: Chang Nie; Zhiyang Ju; Zhifeng Sun; Hui Zhang - **Venue**: *IEEE Transactions on Emerging Topics in Computational Intelligence (TETCI 2023)* - **DOI**: 10.1109/TETCI.2023.3259441 - **Impact & SOTA**: 30 Hz real-time multi-sensor highway perception achieving **86.00% Mostly Tracked (MT)** and **sub-1% (0.92%) Mostly Lost (ML)** on the KITTI benchmark. - **Abstract**: DTFI integrates deep point-cloud and camera feature fusion with an Interactive Multiple Model Unscented Kalman Filter (IMM-UKF) for real-time 3D object detection and trajectory tracking under high-speed highway driving conditions. - **URLs**: - Page: https://changnie.top/publications/dtfi/ - DOI: https://doi.org/10.1109/TETCI.2023.3259441 - **BibTeX**: ```bibtex @article{nie2023dtfi, title={DTFI: LiDAR-Camera Deep Fusion and IMM-UKF Tracking for Highway Perception}, author={Nie, Chang and Ju, Zhiyang and Sun, Zhifeng and Zhang, Hui}, journal={IEEE Transactions on Emerging Topics in Computational Intelligence}, year={2023}, doi={10.1109/TETCI.2023.3259441} } ``` --- ## Flagship Systems & Industrial Deployments ### 1. Multi-Camera SfM 4D Vision Auto-Annotation (NETA Auto) - **Role**: Overall Technical Route Lead & Chief Delivery Engineer - **Impact**: Slashed processing latency by **95%+** (from 30-40 min down to 1-1.5 min), achieved **4.77% dense depth accuracy** against LiDAR ground truth (OEM threshold: 8%), and reduced angular drift by 3x across production sequences. - **Page**: https://changnie.top/projects/sfm/ ### 2. Autonomous Valet Parking (Voyager Tech) - **Role**: Project Lead & Chief Software Architect - **Impact**: Full-stack L4 autonomous valet parking deployed across campus roads and multi-floor underground garages, connecting surround BEV perception, decimeter fusion localization, and Hybrid A* + IPOPT trajectory optimization. - **Page**: https://changnie.top/projects/avp/ ### 3. Commercial Autonomous Lawn Robot (Positec / WORX) - **Role**: System Design Lead & Perception-Fusion Delivery Architect - **Impact**: Commercial lawn mower deployed across non-planar outdoor turf, detecting **7cm micro-trench hazards** and integrating factor-graph RTK/LiDAR/IMU fusion. - **Page**: https://changnie.top/projects/mower/ ### 4. Robotic Chemist for AI for Science - **Role**: Dual-Arm Manipulation Lead & Sim-to-Real Architect - **Impact**: Dual-arm mobile manipulation for chemistry labs, integrating 3D Gaussian Splatting digital twins, millimeter-level labware handling, and centrifuge operations. - **Page**: https://changnie.top/projects/robot-chemist/ ### 5. Industrial Mobile Material Handling Robot - **Role**: Pose Estimation & Dual-Arm Coordination Lead - **Impact**: Heavy-payload mobile manipulation achieving sub-centimeter pallet localization, millimeter 6D pose estimation, and vibration-resistant dual-arm de-palletizing. - **Page**: https://changnie.top/projects/factory-material-handling/ ### 6. Low-Speed Autonomous Research Vehicle Platform - **Role**: Hardware-Software Platform Architect (2019–2022) - **Impact**: Ground-up autonomy stack across low-speed passenger vehicles and 4WIS AGVs, with 8 real-vehicle closed-loop demonstrations. - **Page**: https://changnie.top/projects/lsv/ --- ## Canonical Sitemap & Site Discovery - **English Academic Profile**: https://changnie.top/ - **Chinese Academic Profile (中文主页)**: https://changnie.top/zh/ - **Publications Index**: https://changnie.top/publications/ - **Chinese Publications Index**: https://changnie.top/zh/publications/ - **Projects Index**: https://changnie.top/projects/ - **Chinese Projects Index**: https://changnie.top/zh/projects/ - **XML Sitemap**: https://changnie.top/sitemap.xml - **Concise LLM Guide**: https://changnie.top/llms.txt - **Extended LLM Knowledge Base**: https://changnie.top/llms-full.txt