• Our study, PrimA6D, is accepted for RA-L.
  • Paper, Video
• The extended version, PrimA6D , is accepted for RA-L.
  • Paper, Video

• Method Summary
• Source Code Release for Pose Estimation Method
• Source Code Release for Multi-Object Pose Optimization Method

• PrimA6D (RA-L 2020)

  • PrimA6D reconstructs the rotation primitive and its associated keypoints corresponding to the target object for enhancing the orientation inference.
  
  • More details in PrimA6D: Rotational Primitive Reconstruction for Enhanced and Robust 6D Pose Estimation

• PrimA6D (RA-L 2022)

  • PrimA6D estimates three rotation axis primitive images and their associated uncertainties.
  
  • With estimated uncertainties, PrimA6D handles object ambiguity without prior information on object shape.
  
  • More details in Ambiguity-Aware Multi-Object Pose Optimization for Visually-Assisted Robot Manipulation

• Object-SLAM for Multi-Object Pose Optimization (RA-L 2022)

  • Leveraging the uncertainty, we formulate the problem as an object-SLAM to optimize multi-object poses.
  
  • More details in Ambiguity-Aware Multi-Object Pose Optimization for Visually-Assisted Robot Manipulation

• Download Repo

  $ git clone [email protected]:rpmsnu/PrimA6D.git

• Docker Image Download & Run

  $ docker pull jmong1994/jeon:prima6d_new

  We provide a docker image with an environment setup. You can download this docker image on the docker hub.

• 3D Model Download
  Download 3D Models.

  $ mv 3d_models.zip /path/to/PrimA6D/Pose-Estimation/dataset/3d_model/
  $ cd /path/to/PrimA6D/Pose-Estimation/dataset/3d_model/
  $ unzip 3d_models.zip

• Dataset Download

  • Download Sun2012Pascal and BOP dataset

  $ cd /path/to/PrimA6D/Pose-Estimation/dataset/raw_dataset
  $ bash get_sun2012pascalformat.sh
  $ cd bop          
  $ bash get_bop_ycbv.sh          

• Inference

  • Run docker

  $ xhost  local:docker
  $ docker run --gpus all -it --env="DISPLAY" --net=host --ipc=host --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" -v /:/mydata jmong1994/jeon:prima6d_new bash 
  $ export PrimA6D_path="/path/to/PrimA6D"

  • Prepare data

  $ cd $PrimA6D_path/Pose-Estimation/dataset/YCB
  $ python3 YCB_test.py -o=[obj_id] 

  For example, to prepare the No.1 of YCB object, python3 YCB_train_synthetic.py -o=1

  • Test Model

  $ cd $PrimA6D_path/Pose-Estimation/PrimA6D
  $ python3 4_test_all.py -o=[obj_id] -w         

  For example, to infer the No.1 of YCB object, python3 4_test_all.py -o=1 -w

  For the corresponding object,
  Download PrimA6D weights and save these to $PrimA6D_path/Pose-Estimation/PrimA6D/trained_weight.
  Download Segmentation weights and save these to $PrimA6D_path/Pose-Estimation/Segmentation/trained_weight.

• Train

  • Run docker

  $ xhost  local:docker
  $ docker run --gpus all -it --env="DISPLAY" --net=host --ipc=host --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" -v /:/mydata jmong1994/jeon:prima6d_new bash 
  $ export PrimA6D_path="/path/to/PrimA6D"

  • Prepare data

  $ cd $PrimA6D_path/Pose-Estimation/dataset/YCB
  $ python3 YCB_train_synthetic.py -o=[obj_id]
  $ python3 YCB_train_pbr.py -o=[obj_id]    
  $ python3 YCB_train_real.py -o=[obj_id] 
  $ python3 YCB_test.py -o=[obj_id] 

  For example, to prepare the No.1 of YCB object, python3 YCB_train_synthetic.py -o=1

  • Train & Test model

  $ cd $PrimA6D_path/Pose-Estimation/PrimA6D
  $ python3 1_train_generator.py -o=[obj_id]
  $ python3 2_train_keypoint.py -o=[obj_id]    
  $ python3 3_train_translation.py -o=[obj_id]    
  $ python3 4_test_all.py -o=[obj_id] 

  For example, to train the No.1 of YCB object, python3 1_train_generator.py -o=1

• Inference

  • Run docker

  $ xhost  local:docker
  $ docker run --gpus all -it --env="DISPLAY" --net=host --ipc=host --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" -v /:/mydata jmong1994/jeon:prima6d_new bash 
  $ export PrimA6D_path="/path/to/PrimA6D"

  • Prepare data

  $ cd $PrimA6D_path/Pose-Estimation/dataset/YCB
  $ python3 YCB_test.py -o=[obj_id] 

  For example, to prepare the No.1 of YCB object, python3 YCB_train_synthetic.py -o=1

  • Test model

  $ cd $PrimA6D_path/Pose-Estimation/PrimA6D  
  $ python3 test_prima6d.py -o=[obj_id] -w         

  For example, to infer the No.1 of YCB object, python3 test_prima6d.py -o=1 -w

  For the corresponding object,
  Download PrimA6D weights and save these to $PrimA6D_path/Pose-Estimation/PrimA6D/trained_weight.
  Download Segmentation weights and save these to $PrimA6D_path/Pose-Estimation/Segmentation/trained_weight.

• Train

  • Run docker

  $ xhost  local:docker
  $ docker run --gpus all -it --env="DISPLAY" --net=host --ipc=host --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" -v /:/mydata jmong1994/jeon:prima6d_new bash 
  $ export PrimA6D_path="/path/to/PrimA6D"

  • Prepare data

  $ cd $PrimA6D_path/Pose-Estimation/dataset/YCB
  $ python3 YCB_train_synthetic.py -o=[obj_id]
  $ python3 YCB_train_pbr.py -o=[obj_id]    
  $ python3 YCB_train_real.py -o=[obj_id] 
  $ python3 YCB_test.py -o=[obj_id] 

  For example, to prepare the No.1 of YCB object, python3 YCB_train_synthetic.py -o=1

  • Train & Test model

  $ cd $PrimA6D_path/Pose-Estimation/PrimA6D  
  $ python3 train_prima6d.py -o=[obj_id]
  $ python3 test_prima6d.py -o=[obj_id] 

  For example, to train the No.1 of YCB object, python3 train_prima6d.py -o=1

• Real-Time Demo with ROS

  • Prepare data
    Download simple_demo.bag.

  • Run model

  $ roscore

  In the new terminal,

  $ xhost  local:docker
  $ docker run --gpus all -it --env="DISPLAY" --net=host --ipc=host --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" -v /:/mydata jmong1994/jeon:prima6d_new bash 
  $ export PrimA6D_path="/path/to/PrimA6D"   
  $ cd $PrimA6D_path/Pose-Estimation/ros
  $ python3 ros_PrimD_torch.py -o=[obj_id]         

  For example, to run the No.4 and No.5 of YCB object, python3 ros_PrimD_torch.py -o="4 5"

  In the new terminal,

  $ rosbag play simple_demo.bag           

  • Check result You can check the result using the rviz in the new terminal.

  You can also run this real-time demo using a camera that provides 640x480 RGB data.

• Download Data Processor for file player

  refer to Data Processor

• install osmesa

  $ export PrimA6D_path="/path/to/PrimA6D"
  $ cd $PrimA6D_path/Multi-Object-Pose-Optimization/src/pose_optimization/bop_renderer/osmesa-install/build
  $ sudo mkdir /opt/osmesa
  $ sudo chmod 777 /opt/osmesa
  $ sudo mkdir /opt/llvm
  $ sudo chmod 777 /opt/llvm   
  $ sudo bash ../osmesa-install.sh
  $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/osmesa/lib:/opt/llvm/lib

• build ros package

  $ export PrimA6D_path="/path/to/PrimA6D"
  $ cd $PrimA6D_path/Multi-Object-Pose-Optimization
  $ catkin build --save-config --cmake-args -DCMAKE_BUILD_TYPE=Release

• Downliad 3D models

  • download 3d_model.
  • unzip to $PrimA6D_path/Multi-Object-Pose-Optimization/src/pose_optimization/3d_model/

• run

  $ cd $PrimA6D_path/Multi-Object-Pose-Optimization
  $ source devel/setup.bash
  $ roslaunch pose_optimization run_franka.launch   

  And then,

  Play sensor data using Data Processor

Please consider citing the paper as:

@ARTICLE{jeon-2020-prima6d,
  author={Jeon, Myung-Hwan and Kim, Ayoung},
  journal={IEEE Robotics and Automation Letters}, 
  title={PrimA6D: Rotational Primitive Reconstruction for Enhanced and Robust 6D Pose Estimation}, 
  year={2020},
  volume={5},
  number={3},
  pages={4955-4962},
  doi={10.1109/LRA.2020.3004322}
}

@ARTICLE{jeon-2022-prima6d,
  author={Jeon, Myung-Hwan and Kim, Jeongyun and Ryu, Jee-Hwan and Kim, Ayoung},
  journal={IEEE Robotics and Automation Letters}, 
  title={Ambiguity-Aware Multi-Object Pose Optimization for Visually-Assisted Robot Manipulation}, 
  year={2023},
  volume={8},
  number={1},
  pages={137-144},
  doi={10.1109/LRA.2022.3222998}
}

If you have any questions, contact here please

[email protected]