[0.2.3]Skill Upgrades, Hierarchical Policy Upgrade, Bug Fixes (#959)

* [bugfix] Reconfigure the sim even if the sim is not None

* addressing comments, better way to decide wether or not to reconfigure sim_obj

* same fix for the robot base

* Added back caching to pick task

* Fixed up nav to obj task

* Fixed up open and close skills

* Added back inits for nav to obj

* Batched HRL policy

* Updated task conditions

* Fixed key mismatch in TP-SRL

* Fixed issue with nav to obj skill

* Fixed early stop issue

* Fixed bad filter condition

* Updated training configs

* Refactored and cleaned up config files

* Fixed up reward. Renamed configs

* Added wandb resume

* Refactored eval with debug sensor code.

* Added get policy info

* Hopefully fixed resuming, added safe drop

* Added improved Fetch model

* Removed refactored files

* Added back default

* Removed habitat folder

* Fixed pick task

* Fixed navigation skill

* Removed redundancy in skill training configs

* Addressed some comments

* Removed nav pick

* Update .gitignore

Co-authored-by: Vincent-Pierre BERGES <[email protected]>

* PR comments

* [v0.2.3] Hab2 fixup : merging main and resolving conflicts (#1022)

* [v0.2.3] Add robot document (#1018)

* [v0.2.3] [Stretch] play stretch demo (#1014)

* [Fix] Documentation gym example path (#1016)

Co-authored-by: vincentpierre <[email protected]>

* [0.2.3]Hydra migration 3 v2 (#1000)

* updated md files

* updated habitat2_gym_tutorial

* updated Habitat2_Quickstart

* updated Habitat_Lab

* updated docs

* partially removed YACS (note there are some inpedendent parts of code that still rely on YACS)

* made all DictConfig imports to be from omegaconf using if TYPE_CHECKING

* removed link to the outdated notebook

* fixed tests

* [Hydra Migration] Removing yacs and migratting the dataset generation code of rearrange (#1001)

* removing yacs and migratting the dataset generation code of rearrange

* addressing comments

* typechecking corrections

Co-authored-by: vincentpierre <[email protected]>

* renaming SensorConfig to LabSensorConfig

* fixing svg link

* fixing svg;

* using relative path is better for svg

* addressed comments

* [0.2.3][Hydra Migration] Made agents to be of type Dict[str, AgentConfig] and added migration readme (#1007)

* made habitat.simulator.agents be a Dict[str, AgentConfig]

* addressed comments

* refactored sim agents configuration

* moved rearrange agents declarations to agents folder; finised agents configuration updates; updated docs

* updated get_config import

* Cfg sys migration readme (#2)

* added habitat-lab config readme

* addressed most of the comments

* added main_agent alias

* renamed habitat config reagme and added baselines config readme

* Update habitat-lab/habitat/config/default.py

Co-authored-by: Vincent-Pierre BERGES <[email protected]>

* added desctiption for the agents_order field

* added a dot :)

Co-authored-by: Vincent-Pierre BERGES <[email protected]>

* fixing conflicts

Co-authored-by: Vincent-Pierre BERGES <[email protected]>
Co-authored-by: vincentpierre <[email protected]>

* Fix links in MANIFEST.in (#991)

* Fix broken links in MANIFEST.in

* fixig video_render_view

* fixing issue with gym maker

* editing test that changed with recent config modifications

* [v0.2.3] Fixing mypy (#1023)

* Unpin habitat-lab requirements

* Reorder dependencies

* Fix mypy config

* remove unnecessary mypy args

* Update mypy version

* Fix python lint

* Fix pre-commit mypy further

* fix mypy ignored errors

* fixing mypy

* fixing mypy

* more accurate type:ignores

* adding docs/ to exclude mypy

* trying to fix python_lint

* exlude arguments

* exlude arguments

* Container needs to be a string

Co-authored-by: Aaron Gokaslan <[email protected]>
Co-authored-by: vincentpierre <[email protected]>

* removing the gym envs keys that were removed

Co-authored-by: Jimmy Yang <[email protected]>
Co-authored-by: vincentpierre <[email protected]>
Co-authored-by: Ruslan <[email protected]>
Co-authored-by: Mikaël Dallaire Côté <[email protected]>
Co-authored-by: Aaron Gokaslan <[email protected]>

* Merged

* Removed bad file and fixed IK

* Env fixes

* Fixed precommit

* Fixed eval and fixed wandb logging

* Fix for the resume in VER and the force threshold for open tasks

* Update habitat-baselines/habitat_baselines/rl/ddppo/policy/resnet_policy.py

Co-authored-by: Aaron Gokaslan <[email protected]>

* tp-srl fixes

* tp-srl fixes

* Fixed more tests

* Cleaned up nav action

* Added tests for all tasks in baseline trainer

* Corrected test overrides

* Improved agents loop in resnet_policy

Improved agents loop in resnet_policy

Co-authored-by: Ruslan <[email protected]>

* Fixed policy docs and improvements. Added docs to spawn robot method.

* [0.2.3] 959 PR review suggested changes (#1029)

* removed habitat-baselines/habitat_baselines/config path prefix

* improved adding render view uuids to the gym obs_keys

* made config keys lowercase

* Updated docs

* Update habitat-lab/habitat/tasks/rearrange/rearrange_task.py

Co-authored-by: Vincent-Pierre BERGES <[email protected]>

* Update habitat-baselines/habitat_baselines/rl/ddppo/policy/resnet_policy.py

Co-authored-by: Aaron Gokaslan <[email protected]>

* Fixed formatting

* tp-srl config fix

* fixing documentation ?

Co-authored-by: vincentpierre <[email protected]>
Co-authored-by: ASzot <[email protected]>
Co-authored-by: Vincent-Pierre BERGES <[email protected]>
Co-authored-by: Jimmy Yang <[email protected]>
Co-authored-by: Ruslan <[email protected]>
Co-authored-by: Mikaël Dallaire Côté <[email protected]>
Co-authored-by: Aaron Gokaslan <[email protected]>
Co-authored-by: Aaron Gokaslan <[email protected]>
Co-authored-by: Oleksandr <[email protected]>
Co-authored-by: Oleksandr Maksymets <[email protected]>
Co-authored-by: SergioArnaud <[email protected]>

Author: 9 people

.gitignore

@@ -99,3 +99,4 @@ data
 # outputs from other libraries
 /sandbox
 /plots/outputs
+/wandb/

docs/pages/habitat2.rst

@@ -31,21 +31,10 @@ Habitat includes an implementation of DD-PPO. As an example, start training a pi
 .. code:: sh
 
     python -u habitat_baselines/run.py \
-        --exp-config habitat_baselines/config/rearrange/ddppo_pick.yaml \
+        --exp-config habitat_baselines/config/rearrange/rl_skill.yaml  \
         --run-type train
 
-Find the `complete list of RL configurations here <https://github.com/facebookresearch/habitat-lab/tree/main/habitat-baselines/habitat_baselines/config/rearrange>`__, any config starting with "ddppo" can be substituted into :code:`--exp-config`. See `here <https://github.com/facebookresearch/habitat-lab/tree/main/habitat-baselines/habitat_baselines#baselines>`__  for more information on how to run with Habitat Baselines.
-
-`Home Assistant Benchmark (HAB) Tasks`_
-=======================================
-
-To run the HAB tasks, use any of the training configurations `here <https://github.com/facebookresearch/habitat-lab/tree/main/habitat-baselines/habitat_baselines/config/rearrange/hab>`__. For example, to run monolithic RL training on the Tidy House task run:
-
-.. code:: sh
-
-    python -u habitat_baselines/run.py \
-        --exp-config habitat_baselines/config/rearrange/hab/ddppo_tidy_house.yaml \
-        --run-type train
+This trains the Pick skill by default. To train the other skills, specify: :code:`benchmark/rearrange=skill_name` where :code:`skill_name` can be :code:`close_cab`, :code:`close_fridge`, :code:`open_fridge`, :code:`pick`, :code:`place`, or :code:`nav_to_obj`. See `here <https://github.com/facebookresearch/habitat-lab/tree/main/habitat-baselines/habitat_baselines#baselines>`__  for more information on how to run with Habitat Baselines.
 
 `Task-Planning with Skills RL Baseline`_
 ========================================
@@ -56,20 +45,21 @@ Here we will detail how to run the Task-Planning with Skills trained via reinfor
 .. code:: sh
 
     python -u habitat_baselines/run.py \
-        --exp-config habitat_baselines/config/rearrange/ddppo_place.yaml \
+        --exp-config habitat_baselines/config/rearrange/rl_skill.yaml \
         --run-type train \
-        checkpoint_folder=./place_checkpoints/
+        checkpoint_folder=./place_checkpoints/ \
+        benchmark/rearrange=place
 
 2. To work on HAB tasks, you must also train a :code:`pick`, :code:`nav_to_obj`, :code:`open_cab`, :code:`close_cab`, :code:`open_fridge`, and :code:`close_fridge` policy. To do so, substitute the name of the other skill for :code:`place` in the above command.
 
-3. By default, the TP-SRL baseline will look for the skill checkpoints as :code:`data/models/[skill name].pth` in the Habitat Lab directory as configured `here for each skill <https://github.com/facebookresearch/habitat-lab/blob/710beab2a5500074793b0c8047e3835fdb8f7b7e/habitat_baselines/config/rearrange/hab/tp_srl.yaml#L94>`__. The :code:`tp-srl.yaml` file can be changed to point to the skills you would like to evaluate, or you can copy the model checkpoints in :code:`data/models/`.
+3. By default, the TP-SRL baseline will look for the skill checkpoints as :code:`data/models/[skill name].pth`. The :code:`tp-srl.yaml` file can be changed to point to the skills you would like to evaluate, or you can copy the model checkpoints in :code:`data/models/`.
 
 4. Evaluate the TP-SRL baseline on the :code:`tidy_house` HAB task via:
 
 .. code:: sh
 
     python -u habitat_baselines/run.py \
-        --exp-config habitat_baselines/config/rearrange/hab/tp_srl.yaml \
+        --exp-config habitat_baselines/config/rearrange/tp_srl.yaml \
         --run-type eval \
         benchmark/rearrange=tidy_house
 

examples/interactive_play.py

@@ -604,10 +604,9 @@ def has_pygame():
         help="If true, make the task never end due to reaching max number of steps",
     )
     parser.add_argument(
-        "--add-ik",
+        "--disable-inverse-kinematics",
         action="store_true",
-        default=False,
-        help="If true, changes arm control to IK",
+        help="If specified, does not add the inverse kinematics end-effector control.",
     )
     parser.add_argument(
         "--gfx",
@@ -668,12 +667,12 @@ def has_pygame():
         if args.never_end:
             env_config.max_episode_steps = 0
 
-        if args.add_ik:
+        if not args.disable_inverse_kinematics:
             if "arm_action" not in task_config.actions:
                 raise ValueError(
-                    "Action space does not have any arm control so incompatible with `--add-ik` option"
+                    "Action space does not have any arm control so cannot add inverse kinematics. Specify the `--disable-inverse-kinematics` option"
                 )
-            sim_config.agent_0.ik_arm_urdf = (
+            sim_config.agents.main_agent.ik_arm_urdf = (
                 "./data/robots/hab_fetch/robots/fetch_onlyarm.urdf"
             )
             task_config.actions.arm_action.arm_controller = "ArmEEAction"

examples/tutorials/colabs/Habitat2_Quickstart.ipynb

@@ -477,10 +477,7 @@
     "        lin_speed: 12.0\n",
     "        ang_speed: 12.0\n",
     "        allow_dyn_slide: True\n",
-    "        end_on_stop: False\n",
     "        allow_back: True\n",
-    "        min_abs_lin_speed: 1.0\n",
-    "        min_abs_ang_speed: 1.0\n",
     "  simulator:\n",
     "    type: RearrangeSim-v0\n",
     "    additional_object_paths:\n",

examples/tutorials/colabs/habitat2_gym_tutorial.ipynb

@@ -122,8 +122,6 @@
     "    * `HabitatTidyHouse-v0`\n",
     "    * `HabitatPrepareGroceries-v0`\n",
     "    * `HabitatSetTable-v0`\n",
-    "    * `HabitatNavPick-v0`\n",
-    "    * `HabitatNavPickNavPlace-v0`\n",
     "\n",
     "The Gym environments are automatically registered from the RL training configurations under [\"habitat-lab/habitat/config/benchmark/rearrange\"](https://github.com/facebookresearch/habitat-lab/tree/main/habitat-lab/habitat/config/benchmark/rearrange). The `habitat.gym.auto_name` key in the YAML file determines the `[Task Name]`. The observation keys in `habitat.gym.obs_keys` are what is returned in the observation space. If the the observations are a set of 1D arrays, then the observation space is automatically flattened. For example, in `HabitatReachState-v0` the observation space is `habitat.gym.obs_keys = ['joint', 'relative_resting_position']`. `joint` is a 7D array and `relative_resting_position` is a 3D array. These two arrays are concatenated automatically to give a `10D` observation space. On the other hand, in environments with image observations, the observation is returned as a dictionary.\n",
     "\n",

examples/tutorials/nb_python/Habitat2_Quickstart.py

@@ -426,10 +426,7 @@ class NavPickSuccessMeasurementConfig(MeasurementConfig):
         lin_speed: 12.0
         ang_speed: 12.0
         allow_dyn_slide: True
-        end_on_stop: False
         allow_back: True
-        min_abs_lin_speed: 1.0
-        min_abs_ang_speed: 1.0
   simulator:
     type: RearrangeSim-v0
     additional_object_paths:

examples/tutorials/nb_python/habitat2_gym_tutorial.py

@@ -103,8 +103,6 @@
 #     * `HabitatTidyHouse-v0`
 #     * `HabitatPrepareGroceries-v0`
 #     * `HabitatSetTable-v0`
-#     * `HabitatNavPick-v0`
-#     * `HabitatNavPickNavPlace-v0`
 #
 # The Gym environments are automatically registered from the RL training configurations under ["habitat-lab/habitat/config/benchmark/rearrange"](https://github.com/facebookresearch/habitat-lab/tree/main/habitat-lab/habitat/config/benchmark/rearrange). The `habitat.gym.auto_name` key in the YAML file determines the `[Task Name]`. The observation keys in `habitat.gym.obs_keys` are what is returned in the observation space. If the the observations are a set of 1D arrays, then the observation space is automatically flattened. For example, in `HabitatReachState-v0` the observation space is `habitat.gym.obs_keys = ['joint', 'relative_resting_position']`. `joint` is a 7D array and `relative_resting_position` is a 3D array. These two arrays are concatenated automatically to give a `10D` observation space. On the other hand, in environments with image observations, the observation is returned as a dictionary.
 #

habitat-baselines/habitat_baselines/common/tensorboard_utils.py

@@ -5,10 +5,11 @@
 # LICENSE file in the root directory of this source tree.
 
 import os
-from typing import Any, List
+from typing import Any, List, Optional
 
 import numpy as np
 import torch
+from omegaconf import OmegaConf
 from torch.utils.tensorboard import SummaryWriter
 
 try:
@@ -29,7 +30,13 @@ def get_writer(config, **kwargs):
 
 
 class TensorboardWriter:
-    def __init__(self, log_dir: str, *args: Any, **kwargs: Any):
+    def __init__(
+        self,
+        log_dir: str,
+        *args: Any,
+        resume_run_id: Optional[str] = None,
+        **kwargs: Any,
+    ):
         r"""A Wrapper for tensorboard SummaryWriter. It creates a dummy writer
         when log_dir is empty string or None. It also has functionality that
         generates tb video directly from numpy images.
@@ -44,6 +51,9 @@ def __init__(self, log_dir: str, *args: Any, **kwargs: Any):
         if log_dir is not None and len(log_dir) > 0:
             self.writer = SummaryWriter(log_dir, *args, **kwargs)
 
+    def get_run_id(self) -> Optional[str]:
+        return None
+
     def __getattr__(self, item):
         if self.writer:
             return self.writer.__getattribute__(item)
@@ -90,7 +100,13 @@ def add_video_from_np_images(
 
 
 class WeightsAndBiasesWriter:
-    def __init__(self, config, *args: Any, **kwargs: Any):
+    def __init__(
+        self,
+        config,
+        *args: Any,
+        resume_run_id: Optional[str] = None,
+        **kwargs: Any,
+    ):
         r"""
         Integrates with https://wandb.ai logging service.
         """
@@ -112,9 +128,16 @@ def __init__(self, config, *args: Any, **kwargs: Any):
             raise ValueError(
                 "Requested to log with wandb, but wandb is not installed."
             )
-
-        self.run = wandb.init(
-            config={"slurm": slurm_info_dict, **config}, **wb_kwargs
+        if resume_run_id is not None:
+            wb_kwargs["id"] = resume_run_id
+            wb_kwargs["resume"] = "must"
+
+        self.run = wandb.init(  # type: ignore[attr-defined]
+            config={
+                "slurm": slurm_info_dict,
+                **OmegaConf.to_container(config),  # type: ignore[arg-type]
+            },
+            **wb_kwargs,
         )
 
     def __getattr__(self, item):
@@ -128,14 +151,17 @@ def add_scalars(self, log_group, data_dict, step_id):
             f"{log_group}/{k.replace(' ', '')}": v
             for k, v in data_dict.items()
         }
-        wandb.log(log_data_dict, step=int(step_id))
+        wandb.log(log_data_dict, step=int(step_id))  # type: ignore[attr-defined]
 
     def add_scalar(self, key, value, step_id):
-        wandb.log({key: value}, step=int(step_id))
+        wandb.log({key: value}, step=int(step_id))  # type: ignore[attr-defined]
 
     def __enter__(self):
         return self
 
+    def get_run_id(self) -> Optional[str]:
+        return self.run.id
+
     def __exit__(self, exc_type, exc_val, exc_tb):
         if self.run:
             self.run.finish()