Find What You Want: Learning Demand-conditioned Object Attribute Space for Demand-driven Navigation

This repo is the official implementation of NeurIPS 2023 paper, Demand-driven Navigation

News

An extended version of DDN, Multi-Object Demand-driven Navigation has been accepted as a poster by NeurIPS2024.

TODOs (Under Development):

• README
• Instruction Dataset
• Trajectory Dataset
• Pre-generated Dataset
• Utils Code
• Training
• Testing

Graphic Memory Optimization

update on 2024.11.21: Someone in issue mentioned if it is possible to optimize the graphics memory consumption making it possible to train with 24G of graphics memory. I have made some optimizations to the code, please follow the instructions below.

WARNING: For some personal reasons, I did not try to run this code, just made similar migration changes from my other projects. But I can provide an explanation of the code.

python main.py --epoch=30 --mode=train_DDN_Split --patch_size=25  --workers=32 --dataset_mode=train --device=cuda:0

Some Explanations

In the original --mode=train_DDN, I tried to feed the whole trajectory (maybe 100 steps) into an LSTM to predict the action sequence. To reduce memory consumption, I cut the trajectory into small patches, each of which is at most patch_size. Then I feed the patches into an LSTM to predict the action sequence one patch by one patch. The patch_size is set to 25 in the above command. This may lose some accuracy.

If your GPU memory is not enough, you can try to reduce the patch_size.

Overview

We propose a demand-driven navigation task, which requires an agent to find objects that satisfy human demands, and propose a novel method to solve this task.

Materials Download (Under Updating)

For all dataset and pretrained models, the download link is Googledrive and Onedrive(recommend).

For Chinese, we provide 百度网盘.

Dataset

Instruction Dataset

Please see dataset.

Trajectory Dataset

We provide the raw trajectory data. Please move them to dataset and then unzip them. The following is the structure of the files in the raw_trajectory_dataset.zip package. bc_{train,val}_check.json are the metadata of trajectory dataset.

┌bc
│ ├train
│ │  └house_{idx}
│ │      └path_{idx}
│ │         └{idx}.jpg
│ └val
│    └house_{idx}
│        └path_{idx}
│ │         └{idx}.jpg
├bc_train_check.json
┕bc_val_check.json

Pre-generated Dataset

In order to speed up the training, we use DETR model to segment the image in advance and get the corresponding CLIP-Visual-Feature. It takes $30h$ in a server with dual E5-2680V4 processors and a 22GB RTX 2080Ti graphics card.

python generate_pre_data.py --mode=pre_traj_crop --dataset_mode=train --top_k=16 
python generate_pre_data.py --mode=pre_traj_crop --dataset_mode=val --top_k=16 
python generate_pre_data.py --mode=merge_pre_crop_json 

We have provided the pre-generated dataset in the Materials Download.

Training

Attribute Module

To train the Attribute Module, prepare the following files in the dataset: instruction_{train,val}_check.json, LGO_features.json, instruction_bert_features_check.json

Then run:

python train_attribute_features.py --epoch=5000

Finally, select the model with the lowest loss on the validation set, named attribute_model2.pt.

Navigation Policy

To train the navigation policy, prepare the following files in the dataset: bc_train_{0,1,2,3,4}_pre.h5, bc_{train,val}_check.json, in the pretrained_model: attribute_model2.pt, mae_pretrain_model.pth

Then run

python main.py --epoch=30 --mode=train_DDN  --workers=32 --dataset_mode=train --device=cuda:0

Testing

Model Selection

First, we need to select the model using validation set.

python eval.py --mode=eval_DDN --eval_path=$path_to_saved_model$ --dataset_mode=val  --device=cuda:0 --workers=32

Then we select the model with the highest accuracy on the validation set, assuming its index is $idx$.

Navigation Policy Testing

python eval.py --mode=test_DDN --eval_path=$path_to_saved_model$ --dataset_mode=$train,test$ --seen_instruction=$0,1$  --device=cuda:0 --epoch=500 --eval_ckpt=$idx$

For the parameter dataset_mode, 'train' represents 'seen_scene', while 'test' represents 'unseen_scene'. Just choose one of them during the test.

For the parameter seen_instruction, '1' represents 'seen_instruction', while '0' represents 'unseen_scene'. Just choose one of them during the test.

Note: if you run AI2Thor in a headless machine, xvfb is highly recommended. Here is an example.

xvfb-run -a python eval.py --mode=test_DDN --eval_path=$path_to_saved_model$ --dataset_mode=train --seen_instruction=1  --device=cuda:0 --epoch=500 --eval_ckpt=15

Contact

If you have any suggestion or questions, please feel free to contact us:

Hongcheng Wang: whc.1999@pku.edu.cn

Hao Dong: hao.dong@pku.edu.cn