Multi-Object Navigation with dynamically learned neural implicit representations

Project Page | Paper

This codebase is based on a previous repository studying the impact of auxiliary losses on MultiON performance. With this released code, any user can thus experiment with our introduced implicit representations (see paper) and auxiliary losses from our previous work.

Multi-Object Navigation with dynamically learned neural implicit representations
Pierre Marza¹, Laetitia Matignon², Olivier Simonin¹, Christian Wolf³
¹INSA Lyon, ²Université Lyon 1, ³Naver Labs Europe
in ICCV 2023

Setup

Please follow the instructions in https://github.com/saimwani/multiON to install dependencies and download Matterport3D scenes.

Data

Episodes

Episodes are the same as in our previous work. Download train, val and test episodes here. The 3 folders should be copied to data/datasets/multinav/3_ON/. Please be careful! In the original MultiON codebase, what was called val set referred to the set of data used to perform final test of the model. We re-named their val set into test set, and introduced a proper validation set (val) to perform hyper-parameter search and early stopping.

Pre-trained models

You can download all pre-trained models here. We provide a pre-trained global reader model (global_reader.pth), visual encoder and segmentation head for the Semantic Finder (semantic_finder_visual_encoder.pth and semantic_finder_segmentation_head.pth). Finally, we also release the weights of our SOTA agent (policy.pth) that was trained from scratch without any pre-training or training curriculum and with auxiliary losses, corresponding to row (k) in Table 2 in our paper. In the paper, we report mean and std over different training runs (seeds). We chose the one with highest PPL performance on the validation (val) set.

Agent training

Implicit agent training

You can modify hyperparameters related to implicit representations in habitat_baselines/config/multinav/ppo_multinav.yaml. The two most important are: learn_sem_finder (True/False) to train a Semantic Finder in real time and use its outputs as input to the agent, and learn_impl_net_expl_occ (True/False) to train an Exploration and Occupancy Implicit Representation in real time. If learn_impl_net_expl_occ is True, learn_sem_finder must also be True. Having only learn_sem_finder as True and learn_impl_net_expl_occ as False will lead to faster training (as only one type of implicit representation is trained in real time). Once hyperparameters are set up, you can use the following command to train an agent using implicit representations,

python habitat_baselines/run.py --exp-config habitat_baselines/config/multinav/ppo_multinav.yaml --agent-type implicit --run-type train

Implicit agent training with auxiliary losses

As mentioned, training with auxiliary losses from our previous work is also supported with this code. The following command shows you how to train an agent with implicit representations using the three auiliary losses (seen target, direction prediction, distance prediction) with the same coefficient (0.25) in the final loss,

python habitat_baselines/run.py --exp-config habitat_baselines/config/multinav/ppo_multinav.yaml --agent-type implicit --run-type train --seen_coef 0.25 --dir_coef 0.25 --dist_coef 0.25

Agent evaluation

To evaluate a model simply change the --run-type flag from train to eval and specify the path to either a specific checkpoint, or a folder containing a set of checkpoints (to evaluate all checkpoints in the folder) with the --eval_path flag: in the example below, we evaluate our released SOTA agent on the test set. You can modify the flag value to evaluate any other checkpoint you get after training a model. You can specify the set to evaluate on (val or test) with the --eval_split flag. Be careful not to use more processes than scenes (val and test respectively contain episodes from 11 and 18 scenes).

python habitat_baselines/run.py --exp-config habitat_baselines/config/multinav/ppo_multinav.yaml --agent-type implicit --run-type eval --eval_path pretrained_models/policy.pth --eval_split test

Citation

@inproceedings{marza2023_dynamic_impl_repr,
    author    = {Marza, Pierre and Matignon, Laetitia and Simonin, Olivier and Wolf, Christian},
    title     = {Multi-Object Navigation with Dynamically Learned Neural Implicit Representations},
    booktitle = {International Conference on Computer Vision (ICCV)},
    year      = {2023},
}

PierreMarza / dynamic_implicit_representations