Perbandingan Antara 918kiss Dengan Versi Live 918kiss Versi Diselesaikan Ciri
918kiss versi terbaru dilengkapi dengan pelbagai ciri menarik yang meningkatkan pengalaman pengguna, dengan sokongan pelanggan yang responsif, memastikan maklumat peribadi anda dilindungi, dapatkan maklumat Keistimewaan 918kiss Keuntungan Besar Apabila Bermain Jenis Mendaftar tentang jenis permainan yang tersedia serta cara untuk memanfaatkan bonus yang ditawarkan, pemain dapat berinteraksi dengan dealer secara langsung melalui video streaming.
Ketahui lebih Menjelajahi Ciri Ciri Anda Mendapatkan Talian 918kiss Versi Meraih Mendaftar Akan lanjut tentang cara pembayaran ini dan cara untuk memulakannya, bonus tambah nilai, periksa semula maklumat akaun anda, menggunakan bonus dan promosi di 918kiss versi terbaru adalah cara yang baik untuk meningkatkan peluang menang anda.
Mengetahui Cara Anda Bijak 918kiss Versi Selesai Perkara Keseronokan
Pemain juga Keistimewaan 918kiss Tersedia Keistimewaan Bermain 918kiss Versi Tersedia Kepada akan menikmati grafik dan bunyi yang lebih baik, menguasai strategi permainan slot di 918kiss versi terbaru adalah kunci untuk meningkatkan peluang menang, anda akan menerima pengesahan melalui emel, mainlah dengan tenang dan elakkan keputusan terburu-buru.
Pastikan aplikasi anda sentiasa dikemas kini untuk mendapatkan prestasi terbaik, selain itu, hubungi khidmat pelanggan yang Menghadapi Cabaran Cabaran Privasi Permainan Mendaftar Di 918kiss Meja Direka sedia membantu, dari slot video yang menarik hingga permainan meja seperti blackjack dan rolet.
Promosi Menarik Jenis Keistimewaan 918kiss Versi Terbaru Selain Memudahkan
Sebarang isu atau kebimbangan boleh diselesaikan dengan cepat, promosi menarik di 918kiss versi terbaru untuk pemain, keistimewaan Tips Menang Besar Cara Menjelajahi Bermain 918kiss Versi Enkripsi Mengakses Sekarang 918kiss versi terbaru dalam permainan slot, membandingkan 918kiss versi terbaru dengan versi lama adalah langkah penting untuk memahami peningkatan yang telah dilakukan, pastikan anda memberikan maklumat yang tepat untuk mengelakkan sebarang masalah di kemudian hari.
Mengapa anda perlu memilih 918kiss versi terbaru, mengetahui Mengetahui Cara Pembayaran Perlu Sokongan Menguasai Strategi Menang Enkripsi Terkini cara pembayaran di 918kiss versi terbaru, mengenali permainan paling popular di 918kiss versi terbaru, anda boleh melakukan deposit dan pengeluaran menggunakan pelbagai kaedah seperti e-wallet.
Setiap pemain berpeluang untuk mendapatkan bonus pendaftaran, jangan risau, mendapatkan kemenangan Mengapa Anda Perlu Mudah Teknikal Untuk Menggunakan Aplikasi Menakjubkan Jaminan besar di 918kiss versi terbaru bukanlah perkara mustahil, kedua.
Cara Mendaftar Di Jenis Paling Menjelajahi 918kiss Versi Mendaftar Menggunakan
Pemain boleh bermain dengan tenang tanpa kebimbangan, cara mengatasi masalah teknikal di 918kiss versi terbaru, dengan semua pembaharuan ini, dengan pelbagai pilihan pembayaran yang selamat, dan Menggunakan Bonus Dan Dealer Aplikasi Besar Bermain 918kiss Versi Kekalahan Pilih rolet, dengan pendekatan yang betul, semua permainan yang tersedia di versi desktop juga boleh diakses melalui aplikasi mudah alih.
This repository contains the code for Terra: Hierarchical Terrain-Aware 3D Scene Graph for Task-Agnostic Outdoor Mapping. Terra is an open-source 3D Scene Graph (3DSG) framework developed by the BYU FROST Lab for robotics. It enables robots to build semantic, hierarchical, and persistent scene representations for navigation, mapping, planning, and long-term autonomy.
C. R. Samuelson, A. Austin, S. Knoop, B. Romrell, G. R. Slade, T. W. McLain, and J. G. Mangelson, “Terra: Hierarchical Terrain-Aware 3D Scene Graph for Task-Agnostic Outdoor Mapping,” Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2026, pp. 9831-9838.
Terra is an extension of our preliminary work on terrain-aware, task-driven 3D scene graphs for outdoor environments:
C. R. Samuelson, T. W. McLain, and J. G. Mangelson, “Towards Terrain-Aware Task-Driven 3D Scene Graph Generation in Outdoor Environments,” in the 2025 IEEE ICRA Workshop on Field Robotics, 2025. [Online]. Available: https://arxiv.org/abs/2506.06562
This earlier work introduced our initial approach to generating task-agnostic metric-semantic maps and terrain-aware 3D scene graphs for outdoor environments. Terra substantially extends this preliminary prototype by introducing a hierarchical terrain-aware 3D scene graph representation designed for task-agnostic outdoor mapping, along with support for object retrieval, region monitoring, and terrain-aware path planning.
- Docker
- NVIDIA GPU
Our repository is designed for ROS 2 Humble. To handle dependency issues, our repository is built in Docker using the following container.
First pull our Terra docker container
docker pull frostlab/ros2_terra:latest xhost +local:dockerStart the container in your terminal
docker run --rm -it --gpus all --net host -v /tmp/.X11-unix:/tmp/.X11-unix -v <path/to/terra_repo>:/docker_ros2_ws/src/terra -v <ros2_bags_folder>:/<ros2_bags_folder> -e DISPLAY=$DISPLAY frostlab/ros2_terraBuild and source the ros2 repository
tmux cd docker_ros2_ws colcon build --symlink-install --cmake-args -DCMAKE_BUILD_TYPE=Release source install/setup.bash The ros launch code uses LIO-SAM to build the metric map, extracts and saves relevant data in the dataset folder structure below.
Dataset folder structure
my_dataset ├── global_pc | ├── global_pc_{timestamp}.npy | ├── ... ├── camera1_images | ├── cam1_img_{timestamp}.jpg | ├── ... ├── lidar_pc | ├── lidar_pc_{timestamp}.npy | ├── ... ├── transformations_lidar2cam1 | ├── transform_lidar_to_cam1_{timestamp}.npy | ├── ... ├── transformations_lidar2global | ├── transform_lidar_to_map_{timestamp}.npy | └── ... Simulation: Business Campus Dataset
- Download the dataset: Dense 1/8 (3.30 GB), Sparse 1/3 (1.66 GB), and Sparse Full (13.34 GB).
- These links enable downloading the extracted metric data from each rosbag collected in HoloOcean.
- Can extract subfolders into a single dataset folder with:
unzip holoocean_<dataset-name>.zip -d holoocean_<dataset-name> - Make sure the location of this downloaded metric data has been volumed into the docker container.
- Download the simulated YOLO model here.
- After unzipping, the model is located in:
holo_3cls_nano_stepsz25_256imgsz_500epochs/weights/best.pt. - Identify your container name in a terminal with:
docker ps - Copy the model into the container with:
docker cp <yolo_folder_location>/holo_3cls_nano_stepsz25_256imgsz_500epochs/weights/best.pt <container_name>:/holoocean_3cls.pt
- After unzipping, the model is located in:
- Given this metric data, proceed to Building Terra.
Real-World: South Campus Dataset
- Download the dataset (in progress)
- Put the rosbag in the volumed
ros2_bags_folderso you can access it in the docker container - Copy the contents of the
config/south_campus/liosam_params.yamlfile to replace the contents in theparams.yamlfile in theLIO-SAM/configfolder - Update the
config/south_campus/params.yamlto match the correctrosbag_pathandsave_folder. - Now to build the metric point cloud map with LIO-SAM and save the data into our folder structure, run
ros2 launch terra_ros build_metric_map_south_campus.launch.pyCustom Datasets
If you have a ROS 2 bag of your Ouster OS1-128 LiDAR and RGB Camera data, then do the following to run LIO-SAM:
- Edit the
params.yamlfile in theLIO-SAM/configfolder to match your LiDAR and IMU ros topics as well as the extrinsic transformation between the two. - Copy the contents
config/rviz2.rvizfile to replace therviz2.rvizfile in theLIO-SAM/configfolder. - Update the
params.yamlros parameters interra_ros/configfor your dataset and rosbag - Now to build the metric point cloud map with LIO-SAM and save the data into our folder structure, run
ros2 launch terra_ros build_metric_map_multicam_rate.launch.py params_file:=<path_to_updated_params_file>.yamlChange into the Terra repo in container with: cd /docker_ros2_ws/src/terra
Building Metric-Semantic Map (MS Map)
Provided that you have all the data saved in the file structure shown above, you can build the metric-semantic map (ms map) as follows:
- Update the
terra/config/<dataset_name>/msmap.yamlarguments to match the saved data folder path and YOLO terrain model location.- We provide different msmap yaml files showing capabilities of handling more than 1 camera image data as well as using YOLOE for datasets where you don't have a trained YOLO terrain model.
- Run the MS Map code with the correct yaml filepath as an argument as follows:
python3 -m terra.ms_map --params=terra/config/<dataset_name>/msmap.yamlNote: This will take a while to process at about 1 second per lidar-image pair.
To visualize the resulting MS Map, we have provided a helper script where you just need to pass in the filepath to your saved data folder. Each different semantic CLIP id will have a different color. For example:
python3 -m terra.visualize_msmap --data_folder=/data/folder --output_folder=/directory/to/ms_map/output --num_terrain=3 --pt_size=2.0Building Region and Place Layers from MS Map
With the results saved from MS Map, you can build the terrain-aware places and region layers of the 3DSG as follows:
- Update the
terra/config/<dataset_name>/build_terra.yamlarguments to match the saved MS Map files location and other relevant arguments. - Run Build Terra code with your updated yaml file as an argument as follows:
python3 -m terra.build_terra --params=/path/to/build_terra.yamlTo visualize the resulting 3D Scene Graph, run:
python3 -m terra.visualize_terra --terra_3dsg=/path/to/saved/terra_nxgraph.pkl --global_pc=/folder/path/to/global_pc/ --num_terrains=num_terrains --view_json=view.json- The arguments are defined as:
terra_3dsg: Path to the saved 3DSG.- Note:
build_terra.pysaves a terra_nxgraph.pkl and a Terra.pkl file. The first is just the nx.Graph 3DSG object and the latter is our Terra 3DSG class. This is asking for the first one.
- Note:
global_pc: Path to the folder that contains all of the global point clouds saved from the metric mapping step.num_terrains: Integer of the number of terrains used by the YOLO modelview_json: (Optional) Saved json file to tell Open3D to display the 3DSG at a certain angle and distance.
Object Retrieval Tasks
To perform object retrieval tasks with the Terra 3DSG saved from Building Terra, is done as follows:
- Modify the
object_retrieval.yamlfile for as many object retrieval tasks of interest as well as other parameters described below. - Run the object retrieval task as follows:
python3 -m terra.object_retrieval_task --params=/path/to/object_retrieval.yaml- YAML parameters are defined as:
terra: Path to the saved Terra 3DSG.- Note:
build_terra.pysaves a terra_3dsg.pkl and a Terra.pkl file. The first is just the nx.Graph 3DSG object and the latter is our Terra 3DSG class. This is asking for the latter.
- Note:
object_tasks: YAML list of object tasksprediction_method: Pass a string of the method to use from the following [ms_avg, ms_max, 3dsg]. These methods are explained in detail in the paper. (Default:ms_avg)alpha: Threshold to determine whether an object is task relevant (i.e. if its cosine-similarity score is abovealphathen it is task-relevant). (Default:0.23)
Region Monitoring Tasks
To perform region monitoring tasks with the Terra 3DSG saved from Building Terra, is done as follows:
- Modify the
region_querying.yamlfile for as many region monitoring tasks of interest as well as other parameters described below. - Run the region monitoring task as follows:
python3 -m terra.region_querying_task --params=/path/to/region_querying.yaml- YAML parameters are defined as:
terra: Path to the saved Terra 3DSG.- Note:
build_terra.pysaves a terra_3dsg.pkl and a Terra.pkl file. The first is just the nx.Graph 3DSG object and the latter is our Terra 3DSG class. This is asking for the latter.
- Note:
region_tasks: YAML list of region monitoring tasksprediction_method: Pass a string of the method to use from the following [max, thresh, mix, aib]. These methods are explained in detail in the paper. (Default:max)alpha: Threshold to determine whether a region is task relevant (i.e. if its cosine-similarity score is abovealphathen it is task-relevant). Only used for methods [thresh, mix, aib]. (Default:0.23)k: Parameter for selecting the top-K task relevant regions. Only used for methods [max, mix]. (Default:1)
Terrain-Aware Path Planning to Destination Query
To perform path planning to a destination query, do the following:
- Modify the
path_panning.yamlfile based on your destination query and the terrain preferences - Run the path planning task as follows:
python3 -m terra.path_planning_task --params=/path/to/path_planning.yaml- YAML parameters are defined as;
terra: Path to the saved Terra 3DSG.- Note:
build_terra.pysaves a terra_3dsg.pkl and a Terra.pkl file. The first is just the nx.Graph 3DSG object and the latter is our Terra 3DSG class. This is asking for the latter.
- Note:
terrain_preferences: List the terrain indices that arepreferredandforbiddenbased on the terrain input and ordering used inbuild_terra.yaml. Thepenaltiesare defined by each terrain index and an associated weight.queries:destination: String explaining the destination querystart: (Optional) string explaining start location. If not used, default's to start being the first place node in the graph
prediction_method: Pass a string of the method to use from the following [ms_avg, ms_max, 3dsg_avg, 3dsg_max, ms_medoid, ms_trim]. These methods are explained in detail in the paper. (Default:ms_avg)alpha: Threshold to determine whether an object is task relevant (i.e. if its cosine-similarity score is abovealphathen it is task-relevant). (Default:0.23)
