Robotic Occlusion Reasoning for Efficient Object Existence Prediction

This is the code for our IROS2021 Paper Robotic Occlusion Reasoning for Efficient Object Existence Prediction. The code is mainly (the model part) based on the PyTorch implementation by kevinzakka of Recurrent Models of Visual Attention.

Requirements

• Create and activate the conda environment
  • conda env create -f environment.yml
  • conda activate pyrep
• Install CoppeliaSim
• Install PyRep

Train a model

Train on a specific level of data

Specify the configuration file through --cfg_file:

python code/trainer.py --cfg_file cfg/train-final-model/train-level1-1-vis.yml

Checkpoints and log files are stored in the ./data folder. Visualization by tensorboard:

tensorboard --logdir ./data

Curriculum training

./scripts/start-training.sh

Troubleshooting

• ImportError: libcoppeliaSim.so.1: cannot open shared object file: No such file or directory
  Create a symbolic link named "libcoppeliaSim.so.1" to "libcoppeliaSim.so" manually:
  ln -s /PATH/OF/COPPLELIASIM/libcoppeliaSim.so /PATH/OF/COPPLELIASIM/libcoppeliaSim.so.1

Running Headless

Method 1: Xvfb

This method has no 3D hardware acceleration for rendering but is easy to make a start. To train the model:

xvfb-run python code/trainer.py --cfg_file cfg/train-final-model/train-level1-1-vis.yml

or

xvfb-run ./scripts/start-training.sh

Method 2: Dummy X server

This method provides full 3D hardware acceleration.

• If you have the sudo permission, run sudo python scripts/startx.py. By default, a dummy X server with the display number of "1" will be created. If the display number of "1" is already occupied, you need to assign other free display numbers. Check occupied display numbers by ls /tmp/.X11-unix/.
• If you don't have the sudo permission, run python scripts/generate-startx-conf.py to generate a Xorg configuration file named dummy-xorg.conf by default at the working directory. You need to ask the administrator to copy dummy-xorg.conf to /etc/X11/ and config allowed_users = anybody at /etc/X11/Xwrapper.config. Then you can run Xorg -noreset +extension GLX +extension RANDR +extension RENDER -config dummy-xorg.conf :1 to start a dummy X11 server. :1 is the display number.
• When a dummy X11 server is running, you can train the model by

DISPLAY=:1.0 python code/trainer.py --cfg_file cfg/train-final-model/train-level1-1-vis.yml

or

DISPLAY=:1.0 ./scripts/start-training.sh

By using DISPLAY=:1.0, the first GPU will do the rendering work. If we want to use the second GPU, we should set DISPLAY=:1.1.

Method 3: VirtualGL

Similar to the dummy X server based approach, this method provides full 3D hardware acceleration. Please refer to the README file of PyRep.

Test a model

DISPLAY=:1.0 python code/trainer.py --cfg_file cfg/train-final-model/test.yml

Cite

If you find this work useful, please cite our paper:

@InProceedings{LWKLZLW21,
  author       = "Li, Mengdi and Weber, Cornelius and Kerzel, Matthias and Lee, Jae Hee and Zeng, Zheni and Liu, Zhiyuan and Wermter, Stefan",
  title        = "Robotic Occlusion Reasoning for Efficient Object Existence Prediction",
  booktitle    = "2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)",
  month        = "Oct",
  year         = "2021",
  doi          = "10.1109/IROS51168.2021.9635947",
  url          = "https://www2.informatik.uni-hamburg.de/wtm/publications/2021/LWKLZLW21/LI_IROS2021.pdf"
}

Contact

Mengdi Li - mengdi.li@studium.uni-hamburg.de