Project Page | Paper | Appendix

This is the code for CROWS: Conformalized Reachable Sets for Obstacle Avoidance With Spheres.

CROWS is a real-time, receding horizon trajectory planner that generates probabilitically-safe motion plans based on a neural network-based representation of a spherebased reachable set.

We demonstrate that CROWS outperforms a variety of state-of-the-art methods in solving challenging motion planning tasks in cluttered environments while remaining collision-free.

To set up the Python environment, you can install the required dependencies using conda by following these steps:

conda env create --file environment.yaml
conda activate crows

Note: Solving the conda environment can take up to 15 minutes. If you don't have access to a CUDA-capable device, use the environment-cpu.yaml file to install the CPU-only version of PyTorch.

For MACOS, you will need to modify the cpu environment to remove open3d and then build and install python-fcl and open3d from source manually after activating the environment.

Note that the environment includes:

• zonopy which provides functionalities used for reachability analysis
• zonopy-robots which provides functionalities for specifying and loading robots

MATLAB and CORA 2021 are used to compute Joint Reachable Set in forward_occupancy/jrs_trig/gen_jrs_trig with the provided MATLAB scripts.

The --device argument specifies the device used for PyTorch computations. If a GPU is available, the default setting is --device 0, which uses cuda:0. If only a CPU is available, the default is --device -1, which sets the computation to cpu. For systems with multiple GPUs, you can manually select the device, e.g., --device 2 to use cuda:2.

You can generate the dataset yourself with the following commands:

# Generate dataset for spherical occupancy
python data_processing/generate_dataset.py --batch_size 500  
# Generate dataset for gradients of centers w.r.t. trajectory parameters
python data_processing/generate_dataset.py --batch_size 500 --grad  

You may need to adjust the --batch_size depending on your machine's capacity. A --batch_size of 500 requires approximately 1.4 GB of GPU memory. Increasing the batch size may speed up data generation until you reach a size of 500.

Pretrained models are available in the trained_models/ directory. If you prefer to train your own model, you can run the following commands:

# Train CROWS model for spherical occupancy prediction
python train_model_3d.py --num_epochs 30 --wandb crows --num_workers 5  
# Train CROWS gradient model for center gradients w.r.t. trajectory parameters
python train_model_3d.py --num_epochs 30 --wandb crows --num_workers 5 --train_grad  

You would need to adjust the --num_workers option according to the number of available CPU threads on your machine. The --wandb option uses Weights & Biases for tracking the training process. If you do not want to use it, simply omit the --wandb option.

To run a single trial of CROWS with 20 obstacles and a 0.5-second time limit, use the following command:

python run_statistics_planning_3d.py --planner crows --n_obs 20 --time_limit 0.5 --n_envs 1 --video

The --video argument enables video rendering, and the generated video will be saved in the planning_videos/ directory.

CROWS uses IPOPT, a nonlinear program solver, and the python iterface cyipopt. Experiments in the paper are run with MA27 (ma27) linear solver, which requires to install HSL.

• IPOPT and cyipopt will be install as part of the environment setup of environment.yaml or environment-cpu.yaml.
• For the MA27 (ma27) linear solver, install HSL, seperately. For detailed installation instructions, refer to this guide.
• If the MA27 (ma27) linear solver is unavailable, you can use MUMPS (mumps) but performance may be lower. You can manually specify the linear solver with the --solver option

To reproduce the single-arm planning experiments on random scenarios, run bash run_3d_planning.sh. The results will be in planning_results/ as generated by the planning program.

To reproduce the single-arm planning experiments on hard scenarios, run bash run_scenario_planning.sh. The results will be in scenario_planning_results/ as generated by the planning program.

We generate our figures using a pipeline to export base Blender files which we then add cameras or materials to. Since the Blender code adds quite a bit of bloat, we keep it separate. Please clone the blender-color branch to use those visualization environments.

• zonopy referred some part of CORA.
• This code is built upon sparrows, which provides the core architecture and planning framework that CROWS extends and improves upon.

@inproceedings{kwon2024crows,
  title={Conformalized Reachable Sets for Obstacle Avoidance With Spheres},
  author={Kwon, Yongseok and Michaux, Jonathan and Isaacson, Seth and Zhang, Bohao and Ejakov, Matthew and Skinner, Katherine A and Vasudevan, Ram},
  booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2025},
  organization={IEEE}
}