Merge pull request #1218 from StanfordVL/feat/data-collection

Data collection features and bug fixes

Author: cgokmen

docs/modules/scenes.md

@@ -27,7 +27,7 @@ Every `Environment` instance includes a scene, defined by its config that is pas
       not_load_object_categories: null
       load_room_types: null
       load_room_instances: null
-      seg_map_resolution: 0.1
+      seg_map_resolution: 1.0
     ```
 
 Alternatively, a scene can be directly imported at runtime by first creating the scene class instance (e.g.: `scene = InteractiveTraversableScene(...)`) and then importing it via `og.sim.import_scene(obj)`. This can be useful for iteratively prototyping a desired scene configuration. Note that a scene _must_ be imported before any additional objects are imported!
@@ -49,7 +49,7 @@ The scene keeps track of and organizes all imported objects via its owned `scene
 
 Similarly, systems can be queried via `scene.system_registry`.
 
-In addition, a scene can always be reset by calling `reset()`. The scene's initial state is cached when the scene is first imported, but can manually be updated by calling `scene.update_initial_state(state)`, where `state` can either be a desired state (output of `og.sim.dump_state()`) or `None`, corresponding to the current sim state.
+In addition, a scene can always be reset by calling `reset()`. The scene's initial state is cached when the scene is first imported, but can manually be updated by calling `scene.update_initial_file(scene_file)`, where `scene_file` can either be a desired file (output of `scene.save()`) or `None`, corresponding to the current scene file.
 
 ## Types
 **`OmniGibson`** currently supports two types of scenes. The basic scene class `Scene` implements a minimal scene setup, which can optionally include a skybox and / or ground plane. The second scene class `InteractiveTraversableScene` represents a pre-cached, curated scene exclusively populated with fully-interactive objects from the BEHAVIOR-1K dataset. This scene type additionally includes traversability and semantic maps of the scene floorplan. For a breakdown of all the available scenes and the corresponding objects included in each scene, please refer our [Knowledgebase Dashboard](https://behavior.stanford.edu/knowledgebase/). Below, we provide brief snapshots of each of our 50 BEHAVIOR-1K scenes:

docs/modules/simulator.md

@@ -26,7 +26,7 @@ The simulator can be manually stepped, with or without physics / rendering (`sim
 
 ??? warning annotate "Cycling the sim can cause unexpected behavior"
 
-    If the simulator is playing and then suddenly stopped, all objects are immediately teleported back to their "canonical poses" -- i.e.: the corresponding global poses that were set _before_ the sim started playing. Thus, if `og.sim.play()` is immediately called after, the objects will _not_ teleport back to their respective pre-stopped poses. You can think of the canonical poses as a sort of "initial states", and we recommend explicitly setting a desired initial sim state configuration via `scene.update_initial_state()` and then calling `scene.reset()` after `og.sim.play()`.
+    If the simulator is playing and then suddenly stopped, all objects are immediately teleported back to their "canonical poses" -- i.e.: the corresponding global poses that were set _before_ the sim started playing. Thus, if `og.sim.play()` is immediately called after, the objects will _not_ teleport back to their respective pre-stopped poses. You can think of the canonical poses as a sort of "initial states", and we recommend explicitly setting a desired initial sim scene file configuration via `scene.update_initial_file()` and then calling `scene.reset()` after `og.sim.play()`.
 
 #### Modifying Physics
 If necessary, low-level physics behavior can also be set as well, via the physics interface (`sim.pi`), physics simulation interface (`sim.psi`), physics scene query interface (`sim.psqi`), and physics context (`sim.get_physics_context()`). The simulation timesteps can also be directly set via `sim.set_simulation_dt(...)`.

omnigibson/__init__.py

@@ -5,8 +5,6 @@
 import signal
 import tempfile
 
-import nest_asyncio
-
 from omnigibson.controllers import REGISTERED_CONTROLLERS
 from omnigibson.envs import Environment, VectorEnvironment
 from omnigibson.macros import gm
@@ -25,10 +23,7 @@
     os.getenv("ISAAC_JUPYTER_KERNEL") is not None
 )  # We set this in the kernel.json file
 
-# Always enable nest_asyncio because MaterialPrim calls asyncio.run()
-nest_asyncio.apply()
-
-__version__ = "1.1.1"
+__version__ = "1.2.0-alpha"
 
 root_path = os.path.dirname(os.path.realpath(__file__))
 

omnigibson/action_primitives/starter_semantic_action_primitives.py

@@ -31,6 +31,7 @@
     DifferentialDriveController,
 )
 from omnigibson.macros import create_module_macros
+from omnigibson.macros import macros
 from omnigibson.objects.object_base import BaseObject
 from omnigibson.robots import (
     R1,
@@ -231,6 +232,10 @@ def __init__(
         self._curobo_batch_size = curobo_batch_size
         self.debug_visual_marker = debug_visual_marker
 
+        # Action primitives uses assisted grasping; we ignore delayed AG here
+        with macros.unlocked():
+            macros.robots.manipulation_robot.GRASP_WINDOW = 0.0
+
     @property
     def arm(self):
         assert isinstance(self.robot, ManipulationRobot), "Cannot use arm for non-manipulation robot"

omnigibson/action_primitives/symbolic_semantic_action_primitives.py

@@ -195,7 +195,8 @@ def _release(self):
                 "Cannot release an object if you're not already holding an object",
             )
 
-        self.robot.release_grasp_immediately()
+        for arm in self.robot.arm_names:
+            self.robot.release_grasp_immediately(arm=arm)
         yield from self._settle_robot()
 
     def _toggle(self, obj, value):