Differentiable Planning Library

This is the official codebase for the following two papers:

Integrating Symmetry into Differentiable Planning with Steerable Convolutions

• ICLR 2023
• Linfeng Zhao, Xupeng Zhu, Lingzhi Kong, Robin Walters, Lawson L.S. Wong
• [Paper] [Poster] [Slides] [ICLR page] [OpenReview] [arXiv]

Scaling up and Stabilizing Differentiable Planning with Implicit Differentiation

• ICLR 2023
• Linfeng Zhao, Huazhe Xu, Lawson L.S. Wong
• [Paper] [Poster] [Slides] [ICLR page] [OpenReview] [arXiv]

Papers

Symmetric Planning

Implicit Differentiable Planning

Structure

The codebase is based on a VIN reimplementation pytorch-value-iteration-networks and implementation of GPPN gated-path-planning-networks. It is mostly designed for 2D path planning tasks in the papers.

The helper functions contain the utility library from the Deep Equilibrium Models (DEQ) codebase.

Training Commands

Note: Run all commands under the src directory.

• Step 1: Login to your W&B account

  • Use wandb login <your token>

• Step 2: Generate data

  • Using given maps: 2D navigation

    • python -m envs.generate_dataset --output-path ../data/m15_4abs-cc_10k.npz --mechanism 4abs-cc --maze-size 15 --train-size 10000 --valid-size 2000 --test-size 2000

  • Using given maps: 2-DOF manipulation

    • python -m envs.generate_dataset --env=Arm2DoFsEnv --mechanism=4abs-cc-wrap --maze-size=18 --train-size=10000 --valid-size 2000 --test-size 2000 --max-decimation=0.

  • Using learned maps: 3D visual navigation

    • xvfb-run -a -s "-screen 0 1024x768x24 -ac +extension GLX +render -noreset" python -m envs.generate_dataset --env Visual3DNav --mechanism 4abs-cc --maze-size 15 --train-size 10000 --valid-size 2000 --test-size 2000 --output-path '../data/Visual3DNav_10k_15_4abs-cc'
    • Note: This takes several hours to generate and needs around 10GB storage, and the generation would consume around 20GB memory

  • Using learned maps: 2-DOF manipulation

    • python -m envs.generate_dataset --env=Arm2DoFsWorkSpaceEnv --mechanism=4abs-cc-wrap --maze-size=18 --train-size=10000 --valid-size=1000 --test-size=10000 --max-decimation=0.

• Step 3: Run training

2D navigation. Planning on given maps, without mapper:

python main.py run_train with datafile=../data/m15_4abs-cc_10k.npz mechanism=4abs-cc model=models.SymVIN k=20 f=3 group=d4

Visual navigation. Planning on learned maps, using mapper:

python main.py run_train with datafile=../data/Visual3DNav_10k_15_4abs-cc.npz mechanism=4abs-cc model=models.E2-VIN-v4 k=20 f=3 group=d4 mapper=NavMapper planner_loss=True mapper_loss=True

Contact

Linfeng Zhao (zhao.linf@northeastern.edu)

Citation

@inproceedings{zhao2023integrating,
  title={Integrating Symmetry into Differentiable Planning with Steerable Convolutions},
  author={Zhao, Linfeng and Zhu, Xupeng and Kong, Lingzhi and Walters, Robin and Wong, Lawson LS},
  booktitle={ICLR},
  year={2023}
}

@inproceedings{zhao2023scaling,
  title={Scaling up and Stabilizing Differentiable Planning with Implicit Differentiation},
  author={Zhao, Linfeng and Xu, Huazhe and Wong, Lawson LS},
  booktitle={ICLR},
  year={2023}
}