git submodule update --init --recursive
env.yml
file: conda env create -f env.yml
python -m pip install requirements.txt
python setup.py develop
third_parties
and execute python setup.py develop --all
base_dir
at confs/config.yaml
as the absolute path of this projectAll the data are contained inside the data
directory.
http://dl.fbaipublicfiles.com/habitat/habitat-test-scenes.zip
ln -s /path/to/dataset /path/to/project/data
)python scripts/run_exp.py
run training or deployment of a policy. More information about "RL baselines" and our RL policy below.
To replay an experiment, use the following
python scripts/visualize_exp.py replay.episode_id={ID episode} replay.exp_name={PATH TO EPISODE} replay.modalities="['rgb', 'depth','semantic']"
The following learned baselines are implemented:
neuralslam
: start from confs/habitat/gibson_neuralslam.yaml
seal-v0
: start from confs/habitat/gibson_seal.yaml
curiosity-v0
: start from confs/habitat/gibson_semantic_curiosity.yaml
The following classical baselines are implemented:
randomgoalsbaseline
frontierbaseline-v1
(frontierbaseline-v2
, frontierbaseline-v3
) bouncebaseline
rotatebaseline
randombaseline
Start from confs/habitat/gibson_goal_exploration.yaml
CHECKPOINT_FOLDER
folder in which checkpoints are savedTOTAL_NUM_STEPS
max number of training stepsppo
:
replanning_steps
how often to run the policynum_global_steps
how often to train the policysave_periodic
how often to save a checkpointload_checkpoint_path
full path to a checkpoint to load at startload_checkpoint
set True to load load_checkpoint_path
visualize
if True, debug images are shownEnvironments:
SemanticDisagreement-v0
reward: sum(disagreement_t)Environments for the RL baselines are also provided:
SemanticCuriosity-v0
(Semantic Curiosity)sealenv-v0
(SEAL)ExpSlam-v0
(NeuralSLAM)Policies:
goalexplorationbaseline-v0
State: disagreement_t, map_t, agent poseCheckpoints:
Start from confs/habitat/gibson_goal_exploration.yaml
replanning_steps
how often to run the policyload_checkpoint_path
full path to a checkpoint to load at startload_checkpoint
set to TrueScenes models | Extract path | Archive size |
---|---|---|
Gibson | data/scene_datasets/gibson/{scene}.glb |
1.5 GB |
MatterPort3D | data/scene_datasets/mp3d/{scene}/{scene}.glb |
15 GB |
You can download the task at the following link {ADD LINK}, unzip and put it in data/datasets/objectnav/gibson/v1.1
Task | Scenes | Link | Extract path | Config to use | Archive size |
---|---|---|---|---|---|
Point goal navigation | Gibson | pointnav_gibson_v1.zip | data/datasets/pointnav/gibson/v1/ |
datasets/pointnav/gibson.yaml |
385 MB |
Point goal navigation corresponding to Sim2LoCoBot experiment configuration | Gibson | pointnav_gibson_v2.zip | data/datasets/pointnav/gibson/v2/ |
datasets/pointnav/gibson_v2.yaml |
274 MB |
Point goal navigation | MatterPort3D | pointnav_mp3d_v1.zip | data/datasets/pointnav/mp3d/v1/ |
datasets/pointnav/mp3d.yaml |
400 MB |
https://stanfordvl.github.io/iGibson/dataset.html
wget https://storage.googleapis.com/gibson_scenes/gibson_tiny.tar.gz
If you want to contribute to the project, I suggest to install requirements-dev.txt
and abilitate pre-commit
python -m pip install -r requirements-dev.txt
pre-commit install