Upload folder using huggingface_hub

configs/ddppo_imagenav_gs_eval.yaml

@@ -0,0 +1,99 @@
+# @package _global_
+#
+# DDPPO ImageNav evaluation config for Gaussian Splatting scenes.
+#
+# Usage:
+#   bash scripts_gs/eval_imagenav.sh --ckpt /path/to/checkpoint.pth
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: imagenav
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat_baselines: habitat_baselines_rl_config_base
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: PointNav-v1
+    split: val
+    data_path: data/scene_datasets/gs_scenes/episodes/imagenav/{split}/{split}.json.gz
+    scenes_dir: data/scene_datasets
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 1000
+
+  # ---------- Task ----------
+  task:
+    success_reward: 2.5
+    slack_reward: -1e-3
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/val.scene_dataset_config.json
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 256
+            height: 256
+          depth_sensor:
+            width: 256
+            height: 256
+
+# ---------- Baselines ----------
+habitat_baselines:
+  verbose: False
+  evaluate: True
+  trainer_name: "ddppo"
+  torch_gpu_id: 0
+  tensorboard_dir: "tb"
+  video_dir: "video_dir"
+  test_episode_count: -1
+  num_environments: 1
+  force_torch_single_threaded: True
+  eval_ckpt_path_dir: ""
+  load_resume_state_config: False
+
+  eval:
+    video_option: []
+    use_ckpt_config: False
+
+  rl:
+    policy:
+      main_agent:
+        name: "PointNavResNetPolicy"
+
+    ppo:
+      clip_param: 0.2
+      ppo_epoch: 2
+      num_mini_batch: 2
+      value_loss_coef: 0.5
+      entropy_coef: 0.01
+      lr: 2.5e-4
+      eps: 1e-5
+      max_grad_norm: 0.2
+      num_steps: 64
+      use_gae: True
+      gamma: 0.99
+      tau: 0.95
+      use_linear_clip_decay: False
+      use_linear_lr_decay: False
+      reward_window_size: 50
+      use_normalized_advantage: False
+      hidden_size: 512
+      use_double_buffered_sampler: False
+
+    ddppo:
+      sync_frac: 0.6
+      distrib_backend: NCCL
+      pretrained_weights: ""
+      pretrained: False
+      pretrained_encoder: False
+      train_encoder: True
+      reset_critic: True
+      backbone: resnet50
+      rnn_type: LSTM
+      num_recurrent_layers: 2

configs/ddppo_imagenav_gs_train.yaml

@@ -0,0 +1,100 @@
+# @package _global_
+#
+# DDPPO ImageNav training config for Gaussian Splatting scenes.
+# Uses ImageGoalSensor which renders goal images on-the-fly from the goal position.
+#
+# Usage:
+#   python scripts_gs/run_imagenav.py
+#   bash scripts_gs/train_imagenav.sh --output /path/to/output
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: imagenav
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat_baselines: habitat_baselines_rl_config_base
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: PointNav-v1
+    split: train
+    data_path: data/scene_datasets/gs_scenes/episodes/imagenav/{split}/{split}.json.gz
+    scenes_dir: data/scene_datasets
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 1000
+
+  # ---------- Task ----------
+  task:
+    success_reward: 2.5
+    slack_reward: -1e-3
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/train.scene_dataset_config.json
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 256
+            height: 256
+          depth_sensor:
+            width: 256
+            height: 256
+
+# ---------- Baselines ----------
+habitat_baselines:
+  verbose: False
+  trainer_name: "ddppo"
+  torch_gpu_id: 0
+  tensorboard_dir: "tb"
+  video_dir: "video_dir"
+  test_episode_count: -1
+  eval_ckpt_path_dir: "data/new_checkpoints"
+  num_environments: 4
+  checkpoint_folder: "data/new_checkpoints"
+  num_updates: -1
+  total_num_steps: 2.5e9
+  log_interval: 10
+  num_checkpoints: 100
+  force_torch_single_threaded: True
+
+  rl:
+    policy:
+      main_agent:
+        name: "PointNavResNetPolicy"
+
+    ppo:
+      clip_param: 0.2
+      ppo_epoch: 2
+      num_mini_batch: 2
+      value_loss_coef: 0.5
+      entropy_coef: 0.01
+      lr: 2.5e-4
+      eps: 1e-5
+      max_grad_norm: 0.2
+      num_steps: 64
+      use_gae: True
+      gamma: 0.99
+      tau: 0.95
+      use_linear_clip_decay: False
+      use_linear_lr_decay: False
+      reward_window_size: 50
+      use_normalized_advantage: False
+      hidden_size: 512
+      use_double_buffered_sampler: False
+
+    ddppo:
+      sync_frac: 0.6
+      distrib_backend: NCCL
+      pretrained_weights: ""
+      pretrained: False
+      pretrained_encoder: False
+      train_encoder: True
+      reset_critic: True
+      backbone: resnet50
+      rnn_type: LSTM
+      num_recurrent_layers: 2

configs/ddppo_objectnav_gs_eval.yaml

@@ -0,0 +1,99 @@
+# @package _global_
+#
+# DDPPO ObjectNav evaluation config for Gaussian Splatting scenes.
+#
+# Usage:
+#   bash scripts_gs/eval_objectnav.sh --ckpt /path/to/checkpoint.pth
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: objectnav
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat_baselines: habitat_baselines_rl_config_base
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: ObjectNav-v1
+    split: val
+    data_path: data/scene_datasets/gs_scenes/episodes/objectnav/{split}/{split}.json.gz
+    scenes_dir: data/scene_datasets
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 500
+
+  # ---------- Task ----------
+  task:
+    success_reward: 2.5
+    slack_reward: -1e-3
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/val.scene_dataset_config.json
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 256
+            height: 256
+          depth_sensor:
+            width: 256
+            height: 256
+
+# ---------- Baselines ----------
+habitat_baselines:
+  verbose: False
+  evaluate: True
+  trainer_name: "ddppo"
+  torch_gpu_id: 0
+  tensorboard_dir: "tb"
+  video_dir: "video_dir"
+  test_episode_count: -1
+  num_environments: 1
+  force_torch_single_threaded: True
+  eval_ckpt_path_dir: ""
+  load_resume_state_config: False
+
+  eval:
+    video_option: []
+    use_ckpt_config: False
+
+  rl:
+    policy:
+      main_agent:
+        name: "PointNavResNetPolicy"
+
+    ppo:
+      clip_param: 0.2
+      ppo_epoch: 4
+      num_mini_batch: 2
+      value_loss_coef: 0.5
+      entropy_coef: 0.01
+      lr: 2.5e-4
+      eps: 1e-5
+      max_grad_norm: 0.2
+      num_steps: 64
+      use_gae: True
+      gamma: 0.99
+      tau: 0.95
+      use_linear_clip_decay: False
+      use_linear_lr_decay: False
+      reward_window_size: 50
+      use_normalized_advantage: False
+      hidden_size: 512
+      use_double_buffered_sampler: False
+
+    ddppo:
+      sync_frac: 0.6
+      distrib_backend: NCCL
+      pretrained_weights: ""
+      pretrained: False
+      pretrained_encoder: False
+      train_encoder: True
+      reset_critic: True
+      backbone: resnet50
+      rnn_type: LSTM
+      num_recurrent_layers: 2

configs/ddppo_objectnav_gs_train.yaml

@@ -0,0 +1,100 @@
+# @package _global_
+#
+# DDPPO ObjectNav training config for Gaussian Splatting scenes.
+# Uses SAM+CLIP detected objects with ObjectGoalSensor (category ID lookup).
+#
+# Usage:
+#   python scripts_gs/run_objectnav.py
+#   bash scripts_gs/train_objectnav.sh --output /path/to/output
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: objectnav
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat_baselines: habitat_baselines_rl_config_base
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: ObjectNav-v1
+    split: train
+    data_path: data/scene_datasets/gs_scenes/episodes/objectnav/{split}/{split}.json.gz
+    scenes_dir: data/scene_datasets
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 500
+
+  # ---------- Task ----------
+  task:
+    success_reward: 2.5
+    slack_reward: -1e-3
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/train.scene_dataset_config.json
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 256
+            height: 256
+          depth_sensor:
+            width: 256
+            height: 256
+
+# ---------- Baselines ----------
+habitat_baselines:
+  verbose: False
+  trainer_name: "ddppo"
+  torch_gpu_id: 0
+  tensorboard_dir: "tb"
+  video_dir: "video_dir"
+  test_episode_count: -1
+  eval_ckpt_path_dir: "data/new_checkpoints"
+  num_environments: 4
+  checkpoint_folder: "data/new_checkpoints"
+  num_updates: -1
+  total_num_steps: 2.5e9
+  log_interval: 10
+  num_checkpoints: 100
+  force_torch_single_threaded: True
+
+  rl:
+    policy:
+      main_agent:
+        name: "PointNavResNetPolicy"
+
+    ppo:
+      clip_param: 0.2
+      ppo_epoch: 4
+      num_mini_batch: 2
+      value_loss_coef: 0.5
+      entropy_coef: 0.01
+      lr: 2.5e-4
+      eps: 1e-5
+      max_grad_norm: 0.2
+      num_steps: 64
+      use_gae: True
+      gamma: 0.99
+      tau: 0.95
+      use_linear_clip_decay: False
+      use_linear_lr_decay: False
+      reward_window_size: 50
+      use_normalized_advantage: False
+      hidden_size: 512
+      use_double_buffered_sampler: False
+
+    ddppo:
+      sync_frac: 0.6
+      distrib_backend: NCCL
+      pretrained_weights: ""
+      pretrained: False
+      pretrained_encoder: False
+      train_encoder: True
+      reset_critic: True
+      backbone: resnet50
+      rnn_type: LSTM
+      num_recurrent_layers: 2

configs/ddppo_pointnav_gs_eval.yaml

@@ -0,0 +1,98 @@
+# @package _global_
+#
+# DDPPO PointNav evaluation config for Gaussian Splatting scenes.
+# Self-contained: no extra config files beyond what habitat-lab ships.
+#
+# Usage (from habitat-gs/ root):
+#   python -m habitat_baselines.run \
+#     --config-name=pointnav/ddppo_pointnav_gs_eval
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: pointnav
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat_baselines: habitat_baselines_rl_config_base
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: PointNav-v1
+    split: val
+    data_path: data/scene_datasets/gs_scenes/episodes/pointnav/{split}/{split}.json.gz
+    scenes_dir: data/scene_datasets
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 500
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/val.scene_dataset_config.json
+    # GS scenes supply pre-built navmeshes; do not recompute from collision mesh
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 256
+            height: 256
+          depth_sensor:
+            width: 256
+            height: 256
+
+# ---------- Baselines ----------
+habitat_baselines:
+  verbose: False
+  evaluate: True
+  trainer_name: "ddppo"
+  torch_gpu_id: 0
+  tensorboard_dir: "tb"
+  video_dir: "video_dir"
+  test_episode_count: -1
+  num_environments: 1
+  force_torch_single_threaded: True
+  eval_ckpt_path_dir: ""
+  # Use this config instead of the checkpoint's saved training config
+  load_resume_state_config: False
+
+  eval:
+    video_option: []
+    use_ckpt_config: False
+
+  rl:
+    policy:
+      main_agent:
+        name: "PointNavResNetPolicy"
+
+    ppo:
+      clip_param: 0.2
+      ppo_epoch: 2
+      num_mini_batch: 2
+      value_loss_coef: 0.5
+      entropy_coef: 0.01
+      lr: 2.5e-4
+      eps: 1e-5
+      max_grad_norm: 0.2
+      num_steps: 128
+      use_gae: True
+      gamma: 0.99
+      tau: 0.95
+      use_linear_clip_decay: False
+      use_linear_lr_decay: False
+      reward_window_size: 50
+      use_normalized_advantage: False
+      hidden_size: 512
+      use_double_buffered_sampler: False
+
+    ddppo:
+      sync_frac: 0.6
+      distrib_backend: NCCL
+      pretrained_weights: ""
+      pretrained: False
+      pretrained_encoder: False
+      train_encoder: True
+      reset_critic: True
+      backbone: resnet50
+      rnn_type: LSTM
+      num_recurrent_layers: 2

configs/ddppo_pointnav_gs_train.yaml

@@ -0,0 +1,96 @@
+# @package _global_
+#
+# DDPPO PointNav training config for Gaussian Splatting scenes.
+# Self-contained: no extra config files beyond what habitat-lab ships.
+#
+# Usage (from habitat-gs/ root):
+#   python -m habitat_baselines.run \
+#     --config-name=pointnav/ddppo_pointnav_gs_train
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: pointnav
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat_baselines: habitat_baselines_rl_config_base
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: PointNav-v1
+    split: train
+    data_path: data/scene_datasets/gs_scenes/episodes/pointnav/{split}/{split}.json.gz
+    scenes_dir: data/scene_datasets
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 500
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/train.scene_dataset_config.json
+    # GS scenes supply pre-built navmeshes; do not recompute from collision mesh
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 256
+            height: 256
+          depth_sensor:
+            width: 256
+            height: 256
+
+# ---------- Baselines ----------
+habitat_baselines:
+  verbose: False
+  trainer_name: "ddppo"
+  torch_gpu_id: 0
+  tensorboard_dir: "tb"
+  video_dir: "video_dir"
+  test_episode_count: -1
+  eval_ckpt_path_dir: "data/new_checkpoints"
+  num_environments: 4
+  checkpoint_folder: "data/new_checkpoints"
+  num_updates: -1
+  total_num_steps: 5e8
+  log_interval: 10
+  num_checkpoints: 100
+  force_torch_single_threaded: True
+
+  rl:
+    policy:
+      main_agent:
+        name: "PointNavResNetPolicy"
+
+    ppo:
+      clip_param: 0.2
+      ppo_epoch: 2
+      num_mini_batch: 2
+      value_loss_coef: 0.5
+      entropy_coef: 0.01
+      lr: 2.5e-4
+      eps: 1e-5
+      max_grad_norm: 0.2
+      num_steps: 128
+      use_gae: True
+      gamma: 0.99
+      tau: 0.95
+      use_linear_clip_decay: False
+      use_linear_lr_decay: False
+      reward_window_size: 50
+      use_normalized_advantage: False
+      hidden_size: 512
+      use_double_buffered_sampler: False
+
+    ddppo:
+      sync_frac: 0.6
+      distrib_backend: NCCL
+      pretrained_weights: ""
+      pretrained: False
+      pretrained_encoder: False
+      train_encoder: True
+      reset_critic: True
+      backbone: resnet50
+      rnn_type: LSTM
+      num_recurrent_layers: 2

configs/vln_gs_eval.yaml

@@ -0,0 +1,74 @@
+# @package _global_
+#
+# StreamVLN evaluation config for Gaussian Splatting scenes.
+# Used by StreamVLN evaluation (streamvln_eval.py).
+#
+# Note: StreamVLN training does NOT use habitat configs — it is pure
+# transformer SFT via streamvln_train.py (see scripts_gs/train_vln.sh).
+#
+# Sensor/navigation parameters match StreamVLN defaults:
+#   RGB/Depth: 640x480, hfov=79
+#   forward_step_size=0.25m, turn_angle=15deg
+#   success_distance=3.0m
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: vln_r2r
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgbd_agent
+  - /habitat/task/lab_sensors:
+    - gps_sensor
+    - compass_sensor
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: R2RVLN-v1
+    split: val
+    scenes_dir: data/scene_datasets/
+    data_path: data/scene_datasets/gs_scenes/episodes/vln/{split}/{split}.json.gz
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 500
+    iterator_options:
+      max_scene_repeat_steps: 50000
+      shuffle: False
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/val.scene_dataset_config.json
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 640
+            height: 480
+            hfov: 79
+          depth_sensor:
+            width: 640
+            height: 480
+            hfov: 79
+            min_depth: 0.0
+            max_depth: 10.0
+    forward_step_size: 0.25
+    turn_angle: 15
+    habitat_sim_v0:
+      gpu_device_id: 0
+
+  # ---------- Task / Metrics ----------
+  task:
+    measurements:
+      distance_to_goal:
+        type: DistanceToGoal
+        distance_to: POINT
+      success:
+        type: Success
+        success_distance: 3.0
+      spl:
+        type: SPL
+      oracle_success:
+        type: OracleSuccess
+      oracle_navigation_error:
+        type: OracleNavigationError

configs/vln_uninavid_gs_eval.yaml

@@ -0,0 +1,60 @@
+# @package _global_
+#
+# Uni-NaVid evaluation config for Gaussian Splatting scenes.
+#
+# Sensor/navigation parameters match standard Uni-NaVid:
+#   RGB: 640x480, hfov=120
+#   forward_step_size=0.25m, turn_angle=30deg
+#   success_distance=3.0m
+
+defaults:
+  - /habitat: habitat_config_base
+  - /habitat/task: vln_r2r
+  - /habitat/simulator/sensor_setups@habitat.simulator.agents.main_agent: rgb_agent
+  - /habitat/task/lab_sensors:
+    - gps_sensor
+    - compass_sensor
+  - _self_
+
+# ---------- Dataset ----------
+habitat:
+  dataset:
+    type: R2RVLN-v1
+    split: val
+    scenes_dir: data/scene_datasets/
+    data_path: data/scene_datasets/gs_scenes/episodes/vln/{split}/{split}.json.gz
+
+  # ---------- Environment ----------
+  environment:
+    max_episode_steps: 500
+    iterator_options:
+      max_scene_repeat_steps: 50000
+      shuffle: False
+
+  # ---------- Simulator ----------
+  simulator:
+    scene_dataset: data/scene_datasets/gs_scenes/val.scene_dataset_config.json
+    default_agent_navmesh: False
+    agents:
+      main_agent:
+        sim_sensors:
+          rgb_sensor:
+            width: 640
+            height: 480
+            hfov: 120
+    forward_step_size: 0.25
+    turn_angle: 30
+    habitat_sim_v0:
+      gpu_device_id: 0
+
+  # ---------- Task / Metrics ----------
+  task:
+    measurements:
+      distance_to_goal:
+        type: DistanceToGoal
+        distance_to: POINT
+      success:
+        type: Success
+        success_distance: 3.0
+      spl:
+        type: SPL