Pytorch code for ICRA'21 paper: "Hierarchical Cross-Modal Agent for Robotics Vision-and-Language Navigation"

Overview

Hierarchical Cross-Modal Agent for Robotics Vision-and-Language Navigation

License: MIT PWC

This repository is the pytorch implementation of our paper:

Hierarchical Cross-Modal Agent for Robotics Vision-and-Language Navigation
Muhammad Zubair Irshad, Chih-Yao Ma, Zsolt Kira
International Conference on Robotics and Automation (ICRA), 2021

[Project Page] [arXiv] [GitHub]

Installation

Clone the current repository and required submodules:

git clone https://github.com/GT-RIPL/robo-vln
cd robo-vln
  
export robovln_rootdir=$PWD
    
git submodule init 
git submodule update

Habitat and Other Dependencies

Install robo-vln dependencies as follows:

conda create -n habitat python=3.6 cmake=3.14.0
cd $robovln_rootdir
python -m pip install -r requirements.txt

We use modified versions of Habitat-Sim and Habitat-API to support continuous control/action-spaces in Habitat Simulator. The details regarding continuous action spaces and converting discrete VLN dataset into continuous control formulation can be found in our paper. The specific commits of our modified Habitat-Sim and Habitat-API versions are mentioned below.

# installs both habitat-api and habitat_baselines
cd $robovln_rootdir/environments/habitat-lab
python -m pip install -r requirements.txt
python -m pip install -r habitat_baselines/rl/requirements.txt
python -m pip install -r habitat_baselines/rl/ddppo/requirements.txt
python setup.py develop --all
	
# Install habitat-sim
cd $robovln_rootdir/environments/habitat-sim
python setup.py install --headless --with-cuda

Data

Similar to Habitat-API, we expect a data folder (or symlink) with a particular structure in the top-level directory of this project.

Matterport3D

We utilize Matterport3D (MP3D) photo-realistic scene reconstructions to train and evaluate our agent. A total of 90 Matterport3D scenes are used for robo-vln. Here is the official Matterport3D Dataset download link and associated instructions: project webpage. To download the scenes needed for robo-vln, run the following commands:

# requires running with python 2.7
python download_mp.py --task habitat -o data/scene_datasets/mp3d/

Extract this data to data/scene_datasets/mp3d such that it has the form data/scene_datasets/mp3d/{scene}/{scene}.glb.

Dataset

The Robo-VLN dataset is a continuous control formualtion of the VLN-CE dataset by Krantz et al ported over from Room-to-Room (R2R) dataset created by Anderson et al. The details regarding converting discrete VLN dataset into continuous control formulation can be found in our paper.

Dataset Path to extract Size
robo_vln_v1.zip data/datasets/robo_vln_v1 76.9 MB

Robo-VLN Dataset

The dataset robo_vln_v1 contains the train, val_seen, and val_unseen splits.

  • train: 7739 episodes
  • val_seen: 570 episodes
  • val_unseen: 1224 episodes

Format of {split}.json.gz

{
    'episodes' = [
        {
            'episode_id': 4991,
            'trajectory_id': 3279,
            'scene_id': 'mp3d/JeFG25nYj2p/JeFG25nYj2p.glb',
            'instruction': {
                'instruction_text': 'Walk past the striped area rug...',
                'instruction_tokens': [2384, 1589, 2202, 2118, 133, 1856, 9]
            },
            'start_position': [10.257800102233887, 0.09358400106430054, -2.379739999771118],
            'start_rotation': [0, 0.3332950713608026, 0, 0.9428225683587541],
            'goals': [
                {
                    'position': [3.360340118408203, 0.09358400106430054, 3.07817006111145], 
                    'radius': 3.0
                }
            ],
            'reference_path': [
                [10.257800102233887, 0.09358400106430054, -2.379739999771118], 
                [9.434900283813477, 0.09358400106430054, -1.3061100244522095]
                ...
                [3.360340118408203, 0.09358400106430054, 3.07817006111145],
            ],
            'info': {'geodesic_distance': 9.65537166595459},
        },
        ...
    ],
    'instruction_vocab': [
        'word_list': [..., 'orchids', 'order', 'orient', ...],
        'word2idx_dict': {
            ...,
            'orchids': 1505,
            'order': 1506,
            'orient': 1507,
            ...
        },
        'itos': [..., 'orchids', 'order', 'orient', ...],
        'stoi': {
            ...,
            'orchids': 1505,
            'order': 1506,
            'orient': 1507,
            ...
        },
        'num_vocab': 2504,
        'UNK_INDEX': 1,
        'PAD_INDEX': 0,
    ]
}
  • Format of {split}_gt.json.gz
{
    '4991': {
        'actions': [
          ...
          [-0.999969482421875, 1.0],
          [-0.9999847412109375, 0.15731772780418396],
          ...
          ],
        'forward_steps': 325,
        'locations': [
            [10.257800102233887, 0.09358400106430054, -2.379739999771118],
            [10.257800102233887, 0.09358400106430054, -2.379739999771118],
            ...
            [-12.644463539123535, 0.1518409252166748, 4.2241311073303220]
        ]
    }
    ...
}

Depth Encoder Weights

Similar to VLN-CE, our learning-based models utilizes a depth encoder pretained on a large-scale point-goal navigation task i.e. DDPPO. We utilize depth pretraining by using the DDPPO features from the ResNet50 from the original paper. The pretrained network can be downloaded here. Extract the contents of ddppo-models.zip to data/ddppo-models/{model}.pth.

Training and reproducing results

We use run.py script to train and evaluate all of our baseline models. Use run.py along with a configuration file and a run type (either train or eval) to train or evaluate:

python run.py --exp-config path/to/config.yaml --run-type {train | eval}

For lists of modifiable configuration options, see the default task config and experiment config files.

Evaluating Models

All models can be evaluated using python run.py --exp-config path/to/config.yaml --run-type eval. The relevant config entries for evaluation are:

EVAL_CKPT_PATH_DIR  # path to a checkpoint or a directory of checkpoints
EVAL.USE_CKPT_CONFIG  # if True, use the config saved in the checkpoint file
EVAL.SPLIT  # which dataset split to evaluate on (typically val_seen or val_unseen)
EVAL.EPISODE_COUNT  # how many episodes to evaluate

If EVAL.EPISODE_COUNT is equal to or greater than the number of episodes in the evaluation dataset, all episodes will be evaluated. If EVAL_CKPT_PATH_DIR is a directory, one checkpoint will be evaluated at a time. If there are no more checkpoints to evaluate, the script will poll the directory every few seconds looking for a new one. Each config file listed in the next section is capable of both training and evaluating the model it is accompanied by.

Off-line Data Buffer

All our models require an off-line data buffer for training. To collect the continuous control dataset for both train and val_seen splits, run the following commands before training (Please note that it would take some time on a single GPU to store data. Please also make sure to dedicate around ~1.5 TB of hard-disk space for data collection):

Collect data buffer for train split:

python run.py --exp-config robo_vln_baselines/config/paper_configs/robovln_data_train.yaml --run-type train

Collect data buffer for val_seen split:

python run.py --exp-config robo_vln_baselines/config/paper_configs/robovln_data_val.yaml --run-type train 

CUDA

We use 2 GPUs to train our Hierarchical Model hierarchical_cma.yaml. To train the hierarchical model, dedicate 2 GPUs for training as follows:

CUDA_VISIBLE_DEVICES=0,1 python run.py --exp-config robo_vln_baselines/config/paper_configs/hierarchical_cma.yaml --run-type train

Models/Results From the Paper

Model val_seen SPL val_unseen SPL Config
Seq2Seq 0.34 0.30 seq2seq_robo.yaml
PM 0.27 0.24 seq2seq_robo_pm.yaml
CMA 0.25 0.25 cma.yaml
HCM (Ours) 0.43 0.40 hierarchical_cma.yaml
Legend
Seq2Seq Sequence-to-Sequence. Please see our paper on modification made to the model to match the continuous action spaces in robo-vln
PM Progress monitor
CMA Cross-Modal Attention model. Please see our paper on modification made to the model to match the continuous action spaces in robo-vln
HCM Hierarchical Cross-Modal Agent Module (The proposed hierarchical VLN model from our paper).

Pretrained Model

We provide pretrained model for our best Hierarchical Cross-Modal Agent (HCM). Pre-trained Model can be downloaded as follows:

Pre-trained Model Size
HCM_Agent.pth 691 MB

Citation

If you find this repository useful, please cite our paper:

@inproceedings{irshad2021hierarchical,
title={Hierarchical Cross-Modal Agent for Robotics Vision-and-Language Navigation},
author={Muhammad Zubair Irshad and Chih-Yao Ma and Zsolt Kira},
booktitle={Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
year={2021},
url={https://arxiv.org/abs/2104.10674}
}

Acknowledgments

  • This code is built upon the implementation from VLN-CE
🐤 Nix-TTS: An Incredibly Lightweight End-to-End Text-to-Speech Model via Non End-to-End Distillation

🐤 Nix-TTS An Incredibly Lightweight End-to-End Text-to-Speech Model via Non End-to-End Distillation Rendi Chevi, Radityo Eko Prasojo, Alham Fikri Aji

Rendi Chevi 156 Jan 09, 2023
PyTorch3D is FAIR's library of reusable components for deep learning with 3D data

Introduction PyTorch3D provides efficient, reusable components for 3D Computer Vision research with PyTorch. Key features include: Data structure for

Facebook Research 6.8k Jan 01, 2023
Deep Inside Convolutional Networks - This is a caffe implementation to visualize the learnt model

Deep Inside Convolutional Networks This is a caffe implementation to visualize the learnt model. Part of a class project at Georgia Tech Problem State

Jigar 61 Apr 15, 2022
This is a classifier which basically predicts whether there is a gun law in a state or not, depending on various things like murder rates etc.

Gun-Laws-Classifier This is a classifier which basically predicts whether there is a gun law in a state or not, depending on various things like murde

Awais Saleem 1 Jan 20, 2022
the official implementation of the paper "Isometric Multi-Shape Matching" (CVPR 2021)

Isometric Multi-Shape Matching (IsoMuSh) Paper-CVF | Paper-arXiv | Video | Code Citation If you find our work useful in your research, please consider

Maolin Gao 9 Jul 17, 2022
For IBM Quantum Challenge 2021 (May 20 - 26)

IBM Quantum Challenge 2021 Introduction Commemorating the 40-year anniversary of the Physics of Computation conference, and 5-year anniversary of IBM

Qiskit Community 140 Jan 01, 2023
ThunderGBM: Fast GBDTs and Random Forests on GPUs

Documentations | Installation | Parameters | Python (scikit-learn) interface What's new? ThunderGBM won 2019 Best Paper Award from IEEE Transactions o

Xtra Computing Group 647 Jan 04, 2023
Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields.

This repository contains the code release for Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. This implementation is written in JAX, and is a fork of Google's JaxNeRF

Google 625 Dec 30, 2022
Official implementation of "StyleCariGAN: Caricature Generation via StyleGAN Feature Map Modulation" (SIGGRAPH 2021)

StyleCariGAN: Caricature Generation via StyleGAN Feature Map Modulation This repository contains the official PyTorch implementation of the following

Wonjong Jang 270 Dec 30, 2022
Lepard: Learning Partial point cloud matching in Rigid and Deformable scenes

Lepard: Learning Partial point cloud matching in Rigid and Deformable scenes [Paper] Method overview 4DMatch Benchmark 4DMatch is a benchmark for matc

103 Jan 06, 2023
CDGAN: Cyclic Discriminative Generative Adversarial Networks for Image-to-Image Transformation

CDGAN CDGAN: Cyclic Discriminative Generative Adversarial Networks for Image-to-Image Transformation CDGAN Implementation in PyTorch This is the imple

Kancharagunta Kishan Babu 6 Apr 19, 2022
Machine learning for NeuroImaging in Python

nilearn Nilearn enables approachable and versatile analyses of brain volumes. It provides statistical and machine-learning tools, with instructive doc

919 Dec 25, 2022
The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track.

ISC21-Descriptor-Track-1st The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track. You can check our solution

lyakaap 73 Dec 24, 2022
Implement of "Training deep neural networks via direct loss minimization" in PyTorch for 0-1 loss

This is the implementation of "Training deep neural networks via direct loss minimization" published at ICML 2016 in PyTorch. The implementation targe

Cuong Nguyen 1 Jan 18, 2022
FIRA: Fine-Grained Graph-Based Code Change Representation for Automated Commit Message Generation

FIRA is a learning-based commit message generation approach, which first represents code changes via fine-grained graphs and then learns to generate commit messages automatically.

Van 21 Dec 30, 2022
U^2-Net - Portrait matting This repository explores possibilities of using the original u^2-net model for portrait matting.

U^2-Net - Portrait matting This repository explores possibilities of using the original u^2-net model for portrait matting.

Dennis Bappert 104 Nov 25, 2022
Python Rapid Artificial Intelligence Ab Initio Molecular Dynamics

Python Rapid Artificial Intelligence Ab Initio Molecular Dynamics

14 Nov 06, 2022
RuDOLPH: One Hyper-Modal Transformer can be creative as DALL-E and smart as CLIP

[Paper] [Хабр] [Model Card] [Colab] [Kaggle] RuDOLPH 🦌 🎄 ☃️ One Hyper-Modal Tr

Sber AI 230 Dec 31, 2022
Establishing Strong Baselines for TripClick Health Retrieval; ECIR 2022

TripClick Baselines with Improved Training Data Welcome 🙌 to the hub-repo of our paper: Establishing Strong Baselines for TripClick Health Retrieval

Sebastian Hofstätter 3 Nov 03, 2022
Container : Context Aggregation Network

Container : Context Aggregation Network If you use this code for a paper please cite: @article{gao2021container, title={Container: Context Aggregati

AI2 47 Dec 16, 2022