Dynamic Neural Potential Field: Online Trajectory Optimization in Presence of Moving Obstacles

Official Implementation of Dyn-NPField

Paper | Video | Models | Dataset

OVERVIEW

We address local trajectory planning for a mobile robot in the presence of static and dynamic obstacles. The trajectory is computed as a numerical solution to a Model Predictive Control (MPC) problem, with collision avoidance incorporated by adding obstacle repulsive potential to the MPC cost function. Our approach estimates this repulsive potential using a neural model. We explore three strategies for handling dynamic obstacles: treating them as a sequence of static environments, predicting a full sequence of repulsive potentials at once, and predicting future potentials step by step in an autoregressive mode.

GETTING STARTED

Using Docker

Steps:

1. Clone the Repository: Begin by cloning the repository containing the Dockerfile to your local machine.

   git clone https://github.com/CognitiveAISystems/Dynamic-Neural-Potential-Field
   cd Dynamic-Neural-Potential-Field

2. Build the Docker Image: Use the following command to build the Docker image from the Dockerfile in your repository.

    docker build -t dyn_npfield .

3. Run the Docker Container: After building the image, run a container from it. The following commands will first remove any existing container named dyn_npfield, then start a new one with the specified options. The NPField directory is mounted from your host into the container (instead of being copied):

   docker rm -f dyn_npfield

   docker run -it --gpus all --name dyn_npfield -p 80:80 \
     -v "$(pwd)/NPField:/app/NPField" \
     -v /data/Docker_data/npfiled_dataset:/app/NPField/dataset \
     dyn_npfield

• The -v "$(pwd)/NPField:/app/NPField" bind-mounts your local NPField folder into the container so changes on the host are reflected immediately.
• The -v /data/Docker_data/npfiled_dataset:/app/NPField/dataset mounts a local directory containing your datasets into the container at the given path.

4. Accessing the Container: You can access the running container via:

    docker exec -it dyn_npfield /bin/bash

This will open a bash shell inside the container where you can interact with the software and run commands.

5. Example Run: Inside the container, run the following command to generate the default example GIF:

   export NPFIELD_DATASET_DIR=/app/NPField/dataset/dataset1000
   python NPField/script_d3/NPField_model_GPT.py --finetune-checkpoint /app/NPField/dataset/trained-models/NPField_D3_finetune.pth

This script is intended for testing the neural net and visualizing the resulting neural potential field in the presence of a dynamic obstacle. Parameters for NPField/script_d3/NPField_model_GPT.py:

• episode (int): index into the dataset episode list.
• id_dyn (int): dynamic obstacle id within the episode.
• angle (float, degrees): query heading angle for inference (passed internally in radians). Optional flags:
• --device (cpu or cuda): inference device.
• --chunk-size (int): batch size for grid inference. Output GIFs are written to NPField/output with names like NPField_D3_ep{episode}_dyn{id_dyn}_angle_{angle}deg.gif.

6. Train D3 (new model/checkpoint):

   export NPFIELD_DATASET_DIR=/app/NPField/dataset/dataset1000 && python NPField/script_d3/train_model.py \
     --epochs 10 \
     --lr 5e-5 \
     --batch-size 64 \
     --val-batch-size 16 \
     --n-layer 4 \
     --n-head 4 \
     --n-embd 256 \
     --dropout 0.1 \
     --amp \
     --no-map-loss \
     --checkpoint-name NPField_D3_finetune.pth

7. Test D3 trajectory with the finetuned checkpoint:

   python NPField/script_d3/test_solver_GPT.py \
     --map-id 993 \
     --episodes 10 \
     --finetune-checkpoint /app/NPField/dataset/trained-models/NPField_D3_finetune.pth \
     --save-potential-gif

   python NPField/script_d2/test_solver.py \
     --map-id 993 \
     --episodes 10 \
     --save-potential-gif

8. Generate 100 benchmark configs and Evaluate all scenarios

   # Step 1: 
   cd NPField/config
   python generate_MPC_config.py --save-json --num-scenarios 100

   # Step 2: 
   cd NPField/script_d3
   python test_solver_GPT.py \
      --benchmark-json ../output/benchmark_scenarios.json \
      --finetune-checkpoint /app/NPField/dataset/trained-models/NPField_D3_finetune.pth \
      --save-potential-gif \
      --allow-backward

   cd NPField/script_d2
   python test_solver.py \
      --benchmark-json ../output/benchmark_scenarios.json \
      --save-potential-gif \
      --allow-backward

Troubleshooting:

• Ensure CUDA-compatible drivers are installed on your host machine.
• Check for any Docker-related errors during image building or container running by examining the Docker logs.

CITATION

If you use this framework please cite the following two papers:

Dyn-NPField:

@misc{staroverov2024dynamicneuralpotentialfield,
      title={Dynamic Neural Potential Field: Online Trajectory Optimization in Presence of Moving Obstacles}, 
      author={Aleksey Staroverov and Muhammad Alhaddad and Aditya Narendra and Konstantin Mironov and Aleksandr Panov},
      year={2024},
      eprint={2410.06819},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2410.06819}, 
}

NPField:

@misc{alhaddad2023neuralpotentialfieldobstacleaware,
      title={Neural Potential Field for Obstacle-Aware Local Motion Planning}, 
      author={Muhammad Alhaddad and Konstantin Mironov and Aleksey Staroverov and Aleksandr Panov},
      year={2023},
      eprint={2310.16362},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2310.16362}, 
}