2vin / DECISION

DECISION: Deep rEcurrent Controller for vISual NavigatION

DECISION

DECISION: Deep rEcurrent Controller for vISual NavigatION

This repo contains the implementation of the controllers presented in the paper
Deep Visual Naivgation under Partial Observability, ICRA 2022
This work follows our previous work Intention-Net, CoRL 2017 [paper]

Contents

More specifically, we implemented

• Intention-Net in Torch 1.6.0
• ConvLSTM cells, proposed by the paper
• Truncated Backpropagation Through Time (TBPTT), first appeared in the thesis

Requirements

• PyTorch (ver. 1.6+ required)
• torchvision 0.7.0, tensorboard 2.4.1

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

Refer to environmet.yml for more details of the environment.

Get started

A sample dataset consisting 200 images are available in the repo. To start training our DECISION controller, run the following

git clone https://github.com/AdaCompNUS/DECISION.git
cd DECISION
python main.py

The default hyperparameters may not be optimal. Refer to the below tips for reproduction.

Tips for use

• Do not trust the test loss value. In our case, it is not a good indicator of a control's online performance (i.i.d assumption). Thus, directly evaluate the policy online even for hyperparameter turning. The test value is only a sanity check that the training is ongoing.
• Important hyperparameters: model capacity, learning rate and decay schedule, dropout, reasampling scheme (see below), and arguments --frame-interval, --k1, --k2-n. Refer to the papers for our training hyperparameters.
• One key to leanring a policy that actually works is to adjust the dataset distribution via resampling. Finding the right resampling scheme requires trials and errors. The procedure is dataset-dependent and our implementation is in the SeqDataset class in dataset.py.
• Using a small batch size (> 8 samples per GPU) may bring troubles. If you observe the test loss increases while training loss decreases, try using a larger batch size or commenting out model.eval() before evaluating. If this helps, the problem is the incorrect batch statistics tracked by the norm layers. Solutions: (1) Use a larger batch size. (2) Implement batch norm layers that synchronize statistics across GPUs (nn.SyncBatchNorm might be useful, not tested yet). (3) Use Group Norm (nn.GroupNorm) for all layers and tune the hyperparam num_groups.

Citation

@article{ai2021,
  title={Deep Visual Navigation under Partial Observability},
  author={Ai, Bo and Gao, Wei and Hsu, David and others},
  journal={arXiv preprint arXiv:2109.07752},
  year={2021}
}