SOGMP++/SOGMP: Stochastic Occupancy Grid Map Prediction in Dynamic Scenes

Implementation code for our paper "Stochastic Occupancy Grid Map Prediction in Dynamic Scenes"(arXiv) in Conference on Robot Learning (CoRL) 2023. Two stochastic occupancy grid map (OGM) predictor algorithms (i.e. SOGMP and SOGMP++) implemented by pytorch. Video demos can be found at multimedia demonstrations. Here are three GIFs showing the occupancy grid map prediction comparison results (0.5s, or 5 time steps into the future) of our proposed SOGMP++, SOGMP algorithms, and ConvLSTM, PhyDNet, DeepTracking, and SOGMP_NEMC baselines on three different datasets with different robot models.

• OGM-Turtlebot2:
• OGM-Jackal:
• OGM-Spot:

Requirements

• python 3.7
• torch 1.7.1
• tensorboard

OGM-Datasets

The related datasets can be found at OGM-datasets
There are three different datasets collected by three different robot models (i.e. Turtlebot2, Jackal, Spot).

• 1.OGM-Turtlebot2: collected by a simulated Turtlebot2 with a maximum speed of 0.8 m/s navigates around a lobby Gazebo environment with 34 moving pedestrians using random start points and goal points
• 2.OGM-Jackal: extracted from two sub-datasets of the socially compliant navigation dataset (SCAND), which was collected by the Jackal robot with a maximum speed of 2.0 m/s at the outdoor environment of the UT Austin
• 3.OGM-Spot: extracted from two sub-datasets of the socially compliant navigation dataset (SCAND), which was collected by the Spot robot with a maximum speed of 1.6 m/s at the Union Building of the UT Austin

Usage: SOGMP (The inference speed is faster than SOGMP++)

• Download OGM-datasets and decompress them to the home directory:

  cd ~
  tar -zvxf OGM-datasets.tar.gz

• Training:

  git clone https://github.com/TempleRAIL/SOGMP.git
  cd SOGMP 
  git checkout sogmp
  sh run_train.sh ~/data/OGM-datasets/OGM-Turtlebot2/train ~/data/OGM-datasets/OGM-Turtlebot2/val

• Inference Demo on OGM-Turtlebot2 dataset:

  git clone https://github.com/TempleRAIL/SOGMP.git
  cd SOGMP 
  git checkout sogmp
  sh run_eval_demo.sh  ~/data/OGM-datasets/OGM-Turtlebot2/test

Usage: SOGMP++ (The prediction accuracy is higher than SOGMP)

• Download OGM-datasets and decompress them to the home directory:

  cd ~
  tar -zvxf OGM-datasets.tar.gz

• Training:

  git clone https://github.com/TempleRAIL/SOGMP.git
  cd SOGMP 
  git checkout sogmp++
  sh run_train.sh ~/data/OGM-datasets/OGM-Turtlebot2/train ~/data/OGM-datasets/OGM-Turtlebot2/val

• Inference Demo on OGM-Turtlebot2 dataset:

  git clone https://github.com/TempleRAIL/SOGMP.git
  cd SOGMP 
  git checkout sogmp++
  sh run_eval_demo.sh  ~/data/OGM-datasets/OGM-Turtlebot2/test

Citation

@inproceedings{xie2023stochastic,
  title={Stochastic Occupancy Grid Map Prediction in Dynamic Scenes},
  author={Xie, Zhanteng and Dames, Philip},
  booktitle={Conference on Robot Learning},
  pages={1686--1705},
  year={2023},
  organization={PMLR}
}

@article{xie2023stochastic,
  title={Stochastic Occupancy Grid Map Prediction in Dynamic Scenes},
  author={Xie, Zhanteng and Dames, Philip},
  journal={arXiv preprint arXiv:2210.08577},
  year={2023}
}