cxr0726 / Learning-to-navigate-by-forgetting

The open-sourced code for Learning-to-navigate-by-forgetting

MemoNav: Selecting Informative Memories for Visual Navigation

This repository is the official implementation of MemoNav.

Requirements

The source code is developed and tested in the following setting.

• Python 3.7
• pytorch 1.7.1
• habitat-sim 0.2.0
• habitat 0.2.1

Please refer to habitat-sim and habitat-lab for installation instructions.

To install requirements:

pip install -r requirements.txt

Habitat Data Setup

The scene datasets and task datasets used for training should be organized in the habitat-lab directory as follows:

habitat-api (or habitat-lab)
  └── data
      └── datasets
      │   └── pointnav
      │       └── gibson
      │           └── v1
      │               └── train
      │               └── val
      └── scene_datasets
          └── gibson_habitat
              └── *.glb, *.navmeshs  

The single and multi-goal train/val/test datasets should be organized as follows:

This repo
  └── image-goal-nav-dataset
      |
      └── train
      └── test
      |  └── 1goal
      |  └── 2goal
      |  └── 3goal
      └── val
        └── 1goal
        │   └── *.json.gz
        └── 2goal
        │   └── *.json.gz
        └── 3goal
        │   └── *.json.gz
        └── 4goal
            └── *.json.gz
      

Training

The MemoNav is trained for two phases as the VGM. We first train the agent using imitation learning, minimizing the negative log-likelihood of the ground-truth actions. Next, we finetune the agent’s policy with proximal policy optimization (PPO) to improve the exploratory ability of the agent.

Imitation Learning

To train the model(s) in the paper via IL, run this command:

python train_bc.py --config  ./configs/GATv2_EnvGlobalNode_Respawn_ILRL.yaml --stop

Reinforcement Learning

To fintune the model(s) via RL, run this command:

python train_rl.py --config  ./configs/GATv2_EnvGlobalNode_Respawn_ILRL.yaml --stop --diff hard

Evaluation

To evaluate the model on the single-goal dataset, run:

python evaluate_dataset.py  --config ./configs/GATv2_EnvGlobalNode_Respawn_ILRL.yaml  --eval-ckpt ./data/new_checkpoints/GATv2_EnvGlobalNode_Respawn_ILRL_RL/*.pth --stop --diff hard --gpu 0,0 --forget --version <exp_name>

To evaluate the model on the multi-goal dataset, run:

python evaluate_dataset.py  --config ./configs/GATv2_EnvGlobalNode_Respawn_ILRL.yaml  --eval-ckpt ./data/new_checkpoints/GATv2_EnvGlobalNode_Respawn_ILRL_RL/*.pth --stop --diff 3goal --gpu 0,0 --forget --version <exp_name>

Pre-trained Models

You can download pretrained models here:

• Memonav model trained on Gibson scene datasets.

📋 Give a link to where/how the pretrained models can be downloaded and how they were trained (if applicable). Alternatively you can have an additional column in your results table with a link to the models.

Results

Our model achieves the following performance on:

Gibson single-goal test dataset

Following the experiemntal settings in VGM, our MemoNav model was tested on 1007 samples of this dataset. We reported the performances of our model and various baselines in the table. (NOTE: we re-evaluated the VGM pretrained model and reported new results)

Model name	SR	SPL
ANS	0.30	0.11
Exp4nav	0.47	0.39
SMT	0.56	0.40
Neural Planner	0.42	0.27
NTS	0.43	0.26
VGM	0.75	0.58
MemoNav (ours)	0.78	0.54

Gibson multi-goal test dataset

We compared our model with VGM on multi-goal test datasets which can be downloaded here.

Model name	2goal PR	2goal PPL	3goal PR	3goal PPL	4goal PR	4goal PPL
VGM	0.45	0.18	0.33	0.08	0.28	0.05
MemoNav (ours)	0.50	0.17	0.42	0.09	0.31	0.05

Visualizations

0096_Denmark_success.1.0_spl.0.2_step.218.0.mp4

waypoint_map_0096_Denmark_success.1.0_spl.0.2.mp4

0579_Scioto_success.1.0_spl.0.5_step.137.0.mp4

waypoint_map_0579_Scioto_success.1.0_spl.0.5.mp4

Contributing