Could you share the config file and script to reproduce the ImageNav results for Gibson Split A.

I trained the agent using eval_ppo_imagenav_rgb.yaml (which I'm sure was not intended to be used for training) and got the results:

I can see some discrepancies between the config and the paper (such as not using DDPPO and training hyperparameters).

Thanks!

Hi,
when I run python setup.py install --headless --with-cuda to install habitat-sim, it returns error:

th-cuda
running install
running bdist_egg
running egg_info
writing habitat_sim.egg-info/PKG-INFO
writing dependency_links to habitat_sim.egg-info/dependency_links.txt
writing requirements to habitat_sim.egg-info/requires.txt
writing top-level names to habitat_sim.egg-info/top_level.txt
reading manifest file 'habitat_sim.egg-info/SOURCES.txt'
reading manifest template 'MANIFEST.in'
adding license file 'LICENSE'
writing manifest file 'habitat_sim.egg-info/SOURCES.txt'
installing library code to build/bdist.linux-x86_64/egg
running install_lib
running build_py
running build_ext
error: [Errno 2] No such file or directory: 'build/compile_commands.json'

Is there a method to find bounding box in habit sim?

Hi! When I wanted to run the zer/objectnav dataset , it occurred an error as followed:

File "anaconda3/envs/zsvlnhl/lib/python3.8/site-packages/habitat_sim/simulator.py", line 200, in _config_backend
super().init(config.sim_cfg, config.metadata_mediator)
AssertionError: ESP_CHECK failed: Missing (at least) one of scene dataset attributes, stage attributes, or dataset scene attributes for scene 'data/scene_datasets/gibson_semantic/Markleeville.glb'. Likely an invalid scene name.

But I cannot find the dataset "data/scene_datasets/gibson_semantic" . Can you please inform me the link of the dataset ? Thank you very much!

some config:

habitat:
environment:
type: "gibson"
max_episode_steps: 100

simulator:
turn_angle: 30
tilt_angle: 30
action_space_config: "v1"
agent_0:
sensors: ['rgb_sensor']
height: 0.88
radius: 0.18
habitat_sim_v0:
gpu_device_id: 0
allow_sliding: False
rgb_sensor:
width: 128
height: 128
hfov: 79
position: [0, 0.88, 0]

task:
type: ObjectNav-v1
end_on_success: True
reward_measure: "distance_to_goal_reward"
success_measure: "spl"

possible_actions: ["stop", "move_forward", "turn_left", "turn_right"]

sensors: ['OBJECTgoaltext_sensor', 'compass_sensor', 'gps_sensor']
goal_sensor_uuid: objectgoal

measurements: ['distance_to_goal', 'success', 'spl', 'soft_spl', 'distance_to_goal_reward']

distance_to_goal:
  distance_to: VIEW_POINTS
success:
  success_distance: 0.1

dataset:
type: ObjectNav-v1
split: train
data_path: "data/datasets/zer/objectnav/gibson/v1/{split}/{split}.json.gz"
scenes_dir: "data/scene_datasets/"

habitat_baselines:
base_task_config_path: "exp_config/base_task_config/objectnav_gibson.yaml"
cmd_trailing_opts: ["habitat.environment.iterator_options.max_scene_repeat_steps", "50000"]
simulator_gpu_id: 0
torch_gpu_id: 0
video_option: []
tensorboard_dir: "tb"
video_dir: "video_dir"
test_episode_count: -1
eval_ckpt_path_dir: "data/new_checkpoints"
num_environments: 4
sensors: ["rgb_sensor"]
num_updates: 270000
log_interval: 10
num_checkpoints: 100

Force PyTorch to be single threaded as

this improves performance considerably

force_torch_single_threaded: True

eval:
use_ckpt_config: False
split: val

rl:
policy:
name: "NavNetPolicy"

ppo:
  # para for policy network
  backbone: fast_resnet9
  goal_backbone: clip_text
  rnn_type: GRU
  num_recurrent_layers: 2

  num_steps: 128

  hidden_size: 128
  input_size: 128
  visual_encoder_embedding_size: 512
  goal_embedding_size: 128

  visual_obs_inputs: ['rgb', "imagegoal_sensor_v2"]

  random_crop: False
  rgb_color_jitter: 0.0
  tie_inputs_and_goal_param: False

  task_type_embed: False
  task_type_embed_size: 64

habitat:
simulator:
turn_angle: 30

Hello,

In your ObjectNav for Gibson, is the same dataset used in the ObjectNav Habitat challenge (https://aihabitat.org/challenge/2021/) used? Are the validation tasks themselves same other than the agent height?